A method and apparatus for constructing a prognostic model of pancreatic cancer ferroptosis-related genes

Abstract

This application provides a method and apparatus for constructing a prognostic model for pancreatic cancer ferroptosis-related genes. The method includes: acquiring gene expression datasets and clinical information datasets from pancreatic cancer patients; performing data preprocessing on the gene expression datasets and clinical information datasets to generate a standardized gene expression matrix; performing differential expression analysis and survival analysis based on the standardized gene expression matrix and a ferroptosis pathway gene set to screen out a set of ferroptosis-related characteristic genes significantly associated with prognosis; constructing a deep learning prognostic prediction model based on the ferroptosis-related characteristic gene set; and evaluating the performance of the deep learning prognostic prediction model using an independent validation set to obtain model evaluation metrics. This application can solve the problems of insufficient accuracy and interpretability in existing pancreatic cancer prognostic prediction models.

Description

Technical Field

[0001] This application relates to the field of bioinformatics technology, specifically to a method and apparatus for constructing a prognostic model of pancreatic cancer ferroptosis-related genes. Background Technology

[0002] Pancreatic cancer is a malignant tumor with an extremely poor prognosis, with a 5-year survival rate of less than 10%. Accurate prediction of pancreatic cancer prognosis is crucial for developing personalized treatment plans and improving patient survival. Recent studies have shown that ferroptosis, as a novel programmed cell death mechanism, plays a significant role in tumorigenesis and development. Currently, there are methods for predicting the prognosis of pancreatic cancer using relevant predictive models. However, current pancreatic cancer prognostic prediction models based on ferroptosis-related genes still have the following problems: 1) Existing models mainly rely on single statistical methods, failing to fully explore the complex nonlinear relationships in gene expression data; 2) There is a lack of interpretability analysis of the model prediction results; 3) External validation of the models is insufficient, and their predictive performance across different ethnicities and sequencing platforms has not yet been verified. Summary of the Invention

[0003] In view of this, this application provides a method and apparatus for constructing a prognostic model of pancreatic cancer ferroptosis-related genes based on gene expression data, which solves the problems of insufficient accuracy and interpretability of existing pancreatic cancer prognostic prediction models.

[0004] This application provides a method for constructing a prognostic model of pancreatic cancer ferroptosis-related genes based on gene expression data, including:

[0005] Obtain gene expression datasets and clinical information datasets from pancreatic cancer patients;

[0006] Data preprocessing is performed on the gene expression dataset and the clinical information dataset to generate a standardized gene expression matrix;

[0007] Based on the standardized gene expression matrix, combined with the ferroptosis pathway gene set, differential expression analysis and survival analysis were performed to screen out the ferroptosis characteristic gene set that was significantly associated with prognosis.

[0008] A deep learning prognostic prediction model was constructed based on the aforementioned ferroptosis feature gene set.

[0009] The performance of the deep learning prognostic prediction model was evaluated using an independent validation set to obtain model evaluation metrics.

[0010] Optionally, the gene expression dataset is preprocessed, including:

[0011] Remove gene expression data from the gene expression dataset where the proportion of missing values ​​exceeds 20%;

[0012] The missing values ​​of the remaining gene expression data were filled using the KNN algorithm;

[0013] The ComBat algorithm was used to standardize the filled gene expression data.

[0014] Data preprocessing of the clinical information dataset includes:

[0015] The unit for calculating survival time in the clinical information data of the aforementioned clinical information dataset is unified to months;

[0016] Encode the survival status in the clinical information data;

[0017] Integrate the patient's clinical characteristics information from the aforementioned clinical information data;

[0018] Generate a standardized gene expression matrix, including:

[0019] The processed gene expression data and clinical information data are matched according to the sample ID to generate a standardized gene expression matrix.

[0020] Optionally, the differential expression analysis based on the standardized gene expression matrix and the ferroptosis pathway gene set includes:

[0021] The standardized gene expression matrix is ​​matched with the ferroptosis pathway gene set;

[0022] The expression difference of each matched gene between tumor tissue and normal tissue was calculated to obtain the differential expression analysis results.

[0023] The FDR values ​​of the differential expression analysis results were calculated using the Benjamini-Hochberg method;

[0024] The results of the differential expression analysis were corrected based on the FDR values.

[0025] Optionally, the survival analysis based on the standardized gene expression matrix and the ferroptosis pathway gene set includes:

[0026] Univariate Cox regression analysis was performed on each matched gene to obtain the Cox regression analysis results;

[0027] Based on the Cox regression analysis results, genes that are significantly associated with survival time were screened out;

[0028] Kaplan-Meier survival analysis was performed on genes that were significantly associated with survival time to obtain the survival analysis results.

[0029] Optionally, the set of genes identifying ferroptosis characteristics significantly associated with prognosis includes:

[0030] Based on the results of differential expression analysis and survival analysis, the basic framework of the LASSO regression model was determined.

[0031] Set the candidate range for the penalty parameter λ of the basic framework;

[0032] For the ferroptosis pathway gene set, the optimal λ value was selected using a 10-fold cross-validation method;

[0033] The LASSO regression model was retrained on the ferroptosis pathway gene set using the optimal λ value to obtain the regression coefficient for each gene.

[0034] Features of genes with non-zero regression coefficients are extracted to form a set of ferroptosis characteristic genes.

[0035] Optionally, the construction of a deep learning prognostic prediction model based on the ferroptosis feature gene set includes:

[0036] Design a deep learning model architecture, which includes an input layer, a fully connected layer, a Dropout layer, and an output layer. The dimension of the input layer is the same as the number of genes in the ferroptosis feature gene set.

[0037] By combining the Cox partial likelihood loss function and the ranking loss function, a multi-task learning objective for the deep learning model architecture is defined.

[0038] Add L1 and L2 regularization terms to the multi-task learning objective;

[0039] Based on the multi-task learning objective and the mini-batch stochastic gradient descent algorithm, the deep learning model architecture is trained and optimized to obtain a deep learning prognostic prediction model.

[0040] Optionally, evaluating the performance of the deep learning prognostic prediction model using an independent validation set includes:

[0041] Calculate the model performance metrics of the deep learning prognostic prediction model on an independent validation set, wherein the model performance metrics include: C-index, time-dependent ROC curve, and survival analysis;

[0042] Construct Cox regression and random forest survival models using data from independent validation sets, respectively.

[0043] Calculate the C-index of the Cox regression model on the independent validation set, and the C-index of the random forest survival model on the independent validation set;

[0044] Compare the C-index of the deep learning prognostic prediction model, the C-index of the Cox regression model, and the C-index of the random forest survival model;

[0045] The performance of the deep learning prognostic prediction model was determined based on the comparison results, time-dependent ROC curves, and survival analysis.

[0046] Optionally, the method further includes:

[0047] Multiple external validation sets are obtained, including external validation sets of data from different ethnic groups and external validation sets of data from different sequencing platforms.

[0048] The external prediction performance of the deep learning prognostic prediction model on data from different ethnic groups and different sequencing platforms was evaluated based on multiple external validation sets.

[0049] Based on the external prediction results, the deep learning prognostic prediction model is retrained.

[0050] Optionally, the method further includes:

[0051] Calculate the contribution of each gene to the prediction results;

[0052] Pathway enrichment analysis was performed on high-risk genes whose contribution exceeded the threshold.

[0053] Based on the pathway enrichment analysis results, a personalized prognostic prediction report is generated.

[0054] This application also provides an apparatus for constructing a prognostic model of pancreatic cancer ferroptosis-related genes, comprising:

[0055] The data acquisition module is used to acquire gene expression datasets and clinical information datasets from pancreatic cancer patients.

[0056] The preprocessing module is used to perform data preprocessing on the gene expression dataset and the clinical information dataset to generate a standardized gene expression matrix;

[0057] The feature selection module is used to perform differential expression analysis and survival analysis based on the standardized gene expression matrix and the ferroptosis pathway gene set, and to screen out the ferroptosis feature gene set that is significantly related to prognosis.

[0058] The model building module is used to build a deep learning prognostic prediction model based on the ferroptosis feature gene set.

[0059] The model evaluation module is used to evaluate the performance of the deep learning prognostic prediction model using an independent validation set, and obtain model evaluation metrics.

[0060] This application embodiment also provides a computer device, the computer device comprising:

[0061] At least one processor; and,

[0062] A memory communicatively connected to the at least one processor; wherein,

[0063] The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform the above-described method for constructing a pancreatic cancer ferroptosis-related gene prognostic model based on gene expression data.

[0064] This application also provides a computer-readable storage medium storing computer instructions for causing a computer to execute the above-described method for constructing a pancreatic cancer ferroptosis-related gene prognostic model based on gene expression data.

[0065] This application also provides a computer program product, including computer instructions, characterized in that, when executed by a processor, the computer instructions implement the steps of the above-described method for constructing a prognostic model of pancreatic cancer ferroptosis-related genes based on gene expression data.

[0066] This application has the following technical effects:

[0067] 1. By combining deep learning methods, complex nonlinear relationships in gene expression data can be effectively captured, improving the accuracy of prognostic prediction;

[0068] 2. By analyzing the interpretability of the model's prediction results, key prognostic-related genes and their biological functions can be identified, providing guidance for personalized clinical treatment;

[0069] 3. Validation on multiple external validation sets ensured the model's predictive performance across different ethnicities and sequencing platforms, thus improving the model's generalization ability. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below.

[0071] Figure 1 A flowchart illustrating the method for constructing a pancreatic cancer ferroptosis-related gene prognostic model based on gene expression data, as provided in this application embodiment.

[0072] Figure 2 A schematic diagram of the network structure of the deep learning prognostic prediction model provided in the embodiments of this application.

[0073] Figure 3ROC curve of the model provided in the embodiments of this application on the independent validation set.

[0074] Figure 4 The diagram shows the results of pathway enrichment analysis of high-risk genes provided in the embodiments of this application.

[0075] Figure 5 This is a structural diagram of the apparatus for constructing a pancreatic cancer ferroptosis-related gene prognostic model based on gene expression data, provided in an embodiment of this application. Detailed Implementation

[0076] The embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the embodiments described below with reference to the accompanying drawings are exemplary descriptions for explaining the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions of the embodiments of this application.

[0077] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the terms “comprising” and “including” as used in embodiments of this application mean that the corresponding feature can be implemented as the presented feature, information, data, step, operation, element, and / or component, but do not exclude implementation as other features, information, data, step, operation, element, component, and / or combinations thereof supported by the art. It should be understood that when we say that an element is “connected” or “coupled” to another element, the one element can be directly connected or coupled to the other element, or it can mean that the one element and the other element establish a connection relationship through an intermediate element. Furthermore, “connected” or “coupled” as used herein can include wireless connection or wireless coupling. The term “and / or” as used herein indicates at least one of the items defined by the term; for example, “A and / or B” can be implemented as “A,” or as “B,” or as “A and B.”

[0078] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0079] Example 1

[0080] Reference Figure 1 This application provides a method for constructing a prognostic model of pancreatic cancer ferroptosis-related genes based on gene expression data. This method can accurately predict the prognosis of pancreatic cancer patients, providing a scientific basis for the development of personalized clinical treatment plans. The technical solution of this application will be described in detail below with reference to specific embodiments. The method for constructing a prognostic model of pancreatic cancer ferroptosis-related genes based on gene expression data includes:

[0081] S1: Obtain gene expression datasets and clinical information datasets from pancreatic cancer patients.

[0082] The gene expression dataset contains gene expression profiles of tumor and normal samples, which can be obtained from public databases (such as TCGA, GEO, etc.). The clinical information dataset contains clinical characteristics such as patient survival time, survival status, age, sex, and tumor stage. In this embodiment, RNA sequencing data from 178 pancreatic cancer patients can be obtained from the TCGA-PAAD dataset as a training set, and gene expression data from 69 pancreatic cancer patients can be obtained from the GSE62452 dataset in the GEO database as an independent validation set.

[0083] S2: Perform data preprocessing on the gene expression dataset and the clinical information dataset.

[0084] Preprocessing of gene expression datasets includes the following steps:

[0085] S2.1.1: Perform quality control on gene expression data and remove gene expression data with a missing value ratio exceeding 20%.

[0086] This step can be achieved by calculating the proportion of missing values ​​for each gene across all samples. If the proportion of missing values ​​for a gene exceeds a set threshold, that gene is removed from the dataset. This ensures the data quality for subsequent analyses.

[0087] S2.1.2: For missing values ​​in the remaining data, the K-Nearest Neighbors (KNN) algorithm is used for imputation.

[0088] The KNN algorithm estimates missing values ​​based on the similarity between samples. The specific steps are as follows: First, calculate the Euclidean distance between samples. Then, select the K nearest neighbor samples that are most similar to the missing value sample (K is 10 in this embodiment). Finally, use the average value of the corresponding gene expression values ​​of these K nearest neighbor samples to fill the missing value.

[0089] S2.1.3: Use the ComBat algorithm to perform batch effect elimination and data standardization on the filled gene expression data.

[0090] The ComBat algorithm is a data standardization method based on empirical Bayesian approaches. It can effectively remove systematic differences between data from different batches or platforms. The standardized expression values ​​are typically in the range of [-1, 1], which facilitates subsequent model training.

[0091] S2.2: Preprocessing the clinical information dataset

[0092] Preprocessing clinical information datasets is an important step in ensuring the usability of clinical data, and specifically includes:

[0093] S2.2.1: Standardize the unit of measurement for survival time in clinical information data to months.

[0094] Because different datasets may use different time units (such as days, months, and years) to record patient survival times, unit conversion is necessary. For example, survival time in days is converted to months by dividing by 30.44, and survival time in years is converted to months by multiplying by 12. This standardization process ensures the accuracy of subsequent survival analyses.

[0095] S2.2.2: Encoding the survival state

[0096] Patient survival status is typically encoded in binary format: 0 indicates survival or loss to follow-up, and 1 indicates death. For patients lost to follow-up, their survival time is recorded as the last follow-up time, which conforms to the truncation data processing principle in survival analysis.

[0097] S2.2.3: Integrate the patient's clinical characteristics information

[0098] Integrating patients' clinical characteristics includes: standardizing demographic characteristics such as age and gender; converting TNM staging into a unified AJCC staging standard; encoding categorical variables such as treatment plans and pathological types; deleting clinical characteristics with excessively high missing values; and selecting appropriate imputation strategies for existing missing clinical characteristics based on the characteristic type, such as using the mode to imput categorical variables and the median to imput continuous variables.

[0099] S2.3: Generate a standardized gene expression matrix

[0100] The processed gene expression data and clinical information data are matched according to sample ID to generate a standardized gene expression matrix. The specific implementation process includes:

[0101] First, extract the sample IDs from the gene expression dataset and the clinical information dataset; identify the common sample IDs in both datasets; based on the common sample IDs, extract the corresponding gene expression data from the gene expression dataset; extract the corresponding clinical feature data from the clinical information dataset; merge the gene expression data and clinical feature data according to the sample IDs to form a complete data matrix; check the merged data matrix to ensure the integrity and consistency of the data; standardize the format of the merged data matrix to ensure that the data format meets the needs of subsequent analysis.

[0102] The resulting standardized gene expression matrix has rows representing samples and columns including gene expression values ​​and clinical characteristics. Each element in the matrix has been standardized and is suitable for subsequent analysis.

[0103] S3: Perform differential expression analysis and survival analysis.

[0104] S3 aims to screen for ferroptosis genes associated with pancreatic cancer prognosis from multiple dimensions. Specifically, differential expression analysis and survival analysis were performed, including:

[0105] S3.1: Perform differential expression analysis

[0106] S3.1.1: Match the standardized gene expression matrix with the ferroptosis pathway gene set.

[0107] For example, matching the standardized gene expression matrix with the ferroptosis pathway gene set specifically includes: uniformly converting the gene names in the standardized gene expression matrix into official gene symbols; converting the gene names in the ferroptosis pathway gene set into the same naming system; and using the gene symbols as indexes to locate the expression data of ferroptosis-related genes in the expression matrix.

[0108] S3.1.2: Calculate the expression difference of each gene between tumor tissue and normal tissue.

[0109] For example, calculating the expression difference of each gene between tumor tissue and normal tissue specifically includes: performing differential expression analysis using the limma package; calculating the average expression level difference (logFC value) of each gene; calculating the significance level of the difference (p value) using the modified t-test; and screening differentially expressed genes by considering both the fold change in expression and statistical significance.

[0110] S3.1.3: Calculate FDR values ​​using the Benjamini-Hochberg method

[0111] For example, calculating the FDR value using the Benjamini-Hochberg method specifically includes: sorting all gene p-values ​​in ascending order; calculating the FDR value corresponding to each p-value; and setting an FDR threshold (usually 0.05) for significance assessment.

[0112] S3.1.4: Correct the differential expression analysis results based on the FDR values.

[0113] For example, correcting the differential expression analysis results based on the FDR value specifically includes: screening genes with FDR < 0.05 as significantly differentially expressed genes; visualizing the differential expression analysis results to generate volcano plots and heatmaps; and recording the specific expression change characteristics of each significantly differentially expressed gene.

[0114] S3.2: Constructing a knowledge base of genes related to ferroptosis

[0115] The construction of this knowledge base involved obtaining information from multiple authoritative databases to ensure the completeness and reliability of the gene set. Specifically, the construction of the ferroptosis-related gene set knowledge base included:

[0116] S3.2.1: Extracting genes related to ferroptosis pathways from the KEGG database

[0117] Gene information from pathways such as ferroptosis and iron metabolism can be extracted from the KEGG database. These pathways encompass important biological processes such as iron ion transport, ferritin metabolism, and oxidative stress responses.

[0118] S3.2.2: Obtain iron metabolism pathway genes from the Reactome database

[0119] By accessing the Reactome database, more detailed information on iron metabolism-related pathways can be obtained, including genes involved in processes such as intracellular iron homeostasis regulation and iron-sulfur cluster assembly.

[0120] S3.2.3: Extracting ontology annotations of ferroptosis-related genes from the Gene Ontology database

[0121] The Gene Ontology database allows for the extraction of annotation information on biological processes, molecular functions, and cellular components associated with ferroptosis. This annotation information can help understand the functional properties and cellular localization characteristics of ferroptosis-related genes.

[0122] S3.2.4: Integration of literature-verified ferroptosis genes

[0123] Through literature review, key genes involved in ferroptosis that have been experimentally validated, such as GPX4, SLC7A11, and ACSL4, were integrated. These genes play a central regulatory role in ferroptosis, and their expression levels are closely related to tumor prognosis.

[0124] S3.2.5: Establishing a gene-gene interaction network

[0125] Based on the collected gene information, a gene-gene interaction network was constructed using the String database and biological knowledge. This network not only reflects the physical interactions between genes but also includes functional associations such as transcriptional regulation and signaling pathway cascades. Through network topology analysis, key node genes and functional modules in the ferroptosis pathway can be identified, providing an important basis for subsequent feature selection.

[0126] S3.3: Perform survival analysis

[0127] After completing differential expression analysis and knowledge base construction, survival analysis is needed to assess the association between ferroptosis-related genes and patient prognosis. Specifically, the survival analysis includes:

[0128] S3.3.1: Perform univariate Cox regression analysis

[0129] Univariate Cox proportional hazards regression analysis was performed on each ferroptosis-related gene. This method assesses the strength of the association between individual gene expression levels and patient survival, quantifying this association by calculating the hazard ratio (HR) and the corresponding p-value. For example, an HR value greater than 1 and a p-value less than 0.05 for a gene indicates that high expression of that gene is associated with a poorer prognosis.

[0130] S3.3.2: Screen for genes significantly associated with survival time

[0131] After obtaining the Cox regression analysis results, we focused on those genes that showed statistical significance. The expression levels of these genes were significantly correlated with patient survival and may play an important role in tumor progression.

[0132] S3.3.3: Perform Kaplan-Meier survival analysis

[0133] To further validate the prognostic value of these genes, Kaplan-Meier survival analysis was performed. Patients were divided into high-expression and low-expression groups based on median gene expression levels. Survival curves were plotted, and the log-rank test was used to assess the significance of the survival difference between the two groups. The results of the survival analysis can visually demonstrate the relationship between gene expression levels and patient prognosis, and provide important biological evidence for subsequent model construction.

[0134] By combining the results of differential expression analysis and survival analysis, the potential value of each ferroptosis-related gene in predicting pancreatic cancer prognosis can be comprehensively assessed. This multidimensional analytical strategy not only considers changes in gene expression levels but also focuses on the actual impact of these changes on patient prognosis, thus providing a reliable scientific basis for subsequent feature selection.

[0135] S4: Screening for prognostic ferroptosis-related gene sets

[0136] Specifically, the gene set for screening prognostic-related ferroptosis features includes:

[0137] S4.1: Determine the basic framework of the LASSO regression model

[0138] To screen for the most valuable prognostic feature set from numerous ferroptosis-related genes, this application employs the LASSO (Least Absolute Shrinkage and Selection Operator) regression method for feature selection. This method, by introducing an L1 regularization term, achieves model sparsity during feature selection, avoiding overfitting. The basic framework of the LASSO regression model is based on the Cox proportional hazards model, using patient survival time as the dependent variable and the expression values ​​of previously screened ferroptosis-related genes as independent variables.

[0139] S4.2: Set the candidate range for the penalty parameter λ

[0140] In the model, a penalty term λ is introduced to control the strictness of feature selection. The candidate range of the penalty parameter λ is typically set to a logarithmic sequence, for example, from 10... -4 Up to 10 1 100 values ​​are selected at equal intervals as candidates.

[0141] S4.3: Use 10-fold cross-validation to select the optimal λ value

[0142] To select the optimal λ value, a 10-fold cross-validation method is employed. Specifically, the training set is randomly divided into 10 subsets of similar size. Nine subsets are used to train the model each time, with the remaining subset used for validation. By calculating the cross-validation error of the model under different λ values ​​(typically using partial likelihood bias as an evaluation metric), the λ value that minimizes the cross-validation error is selected as the optimal parameter.

[0143] S4.4: Retrain the LASSO regression model using the optimal λ value

[0144] By retraining the LASSO regression model using the optimal λ value, the regression coefficients for each gene can be obtained. Due to the characteristics of LASSO, the regression coefficients for some genes will be compressed to 0, and these genes will be excluded from the final feature set.

[0145] S4.5: Extract gene features with non-zero regression coefficients

[0146] Genes with non-zero regression coefficients were extracted to form a set of ferroptosis feature genes. These genes are not only significantly associated with patient prognosis but also have strong predictive power, and can be used as input features for subsequent deep learning models.

[0147] In this embodiment, the feature gene set screened by LASSO regression typically contains 10-30 genes. The expression patterns of these genes can effectively reflect the patient's prognostic risk and also facilitate subsequent model construction and clinical application. This feature selection method ensures both the predictive performance of the model and improves its interpretability.

[0148] S5: Constructing a deep learning prognostic prediction model

[0149] Specifically, constructing a deep learning prognostic prediction model includes:

[0150] S5.1: Designing Deep Learning Model Architecture

[0151] like Figure 2 After obtaining the gene set characteristic of ferroptosis, this application employs deep learning methods to construct a prognostic prediction model. The design of the deep learning model fully considers the characteristics of gene expression data and the requirements of the prognostic prediction task, using a multi-layer neural network structure to capture the complex nonlinear relationships between genes.

[0152] The deep learning model's architecture employs a multi-layer feedforward neural network structure. The input layer has the same dimension as the number of genes in the ferroptosis feature gene set, with each neuron corresponding to the expression value of one gene. Following the input layer, multiple fully connected layers are used for feature transformation and information extraction. To prevent overfitting, Dropout layers are added between adjacent fully connected layers, with a dropout rate of 0.3. The final layer of the network is the output layer, used to predict the patient's prognostic risk score. The entire network uses ReLU as the activation function; this choice effectively alleviates the vanishing gradient problem and accelerates model convergence.

[0153] S5.2: Define multi-task learning objectives

[0154] The model's learning objective employs a multi-task learning framework, considering both Cox partial likelihood loss and ranking loss. The Cox partial likelihood loss describes the relationship between patient survival time and predicted risk, while the ranking loss ensures the model can correctly distinguish between high-risk and low-risk patients. Specifically, for any two samples, if one sample has a shorter actual survival time, the model should predict a higher risk score for that sample. This multi-task learning strategy improves the model's predictive accuracy and generalization ability.

[0155] S5.3: Add L1 and L2 regularization terms

[0156] To further improve the model's generalization ability, L1 and L2 regularization terms were added to the loss function. L1 regularization enables feature selection, while L2 regularization prevents the model from over-relying on certain features. The weights of the two regularization terms were determined through cross-validation. The model was trained using a mini-batch stochastic gradient descent algorithm with a batch size of 32, an initial learning rate of 0.001, and a learning rate decay strategy.

[0157] S5.4: Use mini-batch stochastic gradient descent algorithm for training and optimization

[0158] Mini-batch stochastic gradient descent (SGD) is an algorithm used to optimize machine learning models, particularly neural networks. It's a variant of stochastic gradient descent (SGD) designed to balance computational efficiency and convergence speed. In SGD, each parameter update is based on only one training sample (or one data point). This means the algorithm updates the model's weights very frequently, which can lead to instability during training and potentially oscillate around local minima instead of converging to the global minimum. Mini-batch SGD addresses this by using a small batch of training samples for each parameter update. This small batch typically contains tens to hundreds of samples.

[0159] During training, an early stopping strategy can be employed to prevent overfitting. Specifically, the loss function value on the validation set is monitored, and training is stopped when the loss on the validation set fails to improve over several consecutive epochs. This allows training to terminate promptly when the model reaches optimal performance, avoiding overfitting to the training data. This design ensures that the deep learning model can effectively learn prognostic-related feature patterns from gene expression data.

[0160] S6: Evaluate model performance

[0161] Evaluation of deep learning prognostic models is a crucial step in verifying their effectiveness. This application employs multiple evaluation metrics and validation methods to comprehensively evaluate the model's predictive performance. Specifically, the evaluation of model performance includes:

[0162] S6.1: Performance Indicators of the Computational Model

[0163] On the independent validation set, the C-index (consistency index) of the model is first calculated. The C-index reflects the accuracy of the model in ranking patients' prognostic risks, and its value ranges from 0.5 to 1. The higher the value, the stronger the model's discriminative ability. Good prognostic prediction models can usually achieve a C-index value of 0.7 or higher.

[0164] To evaluate the predictive accuracy of the model at different time points, the time-dependent ROC curve and its area under the curve (AUC value) are calculated, such as... Figure 3 As shown. Typically, three time points—1 year, 3 years, and 5 years—are selected to analyze the model's short-term, medium-term, and long-term predictive capabilities.

[0165] Furthermore, the risk stratification effect of the model was evaluated through survival analysis. For example, based on the risk scores predicted by the model, patients were divided into high-risk and low-risk groups, survival curves were plotted using the Kaplan-Meier method, and the statistical significance of the survival difference between the two groups was assessed using the log-rank test.

[0166] Furthermore, the robustness of a model can be assessed by analyzing its predictive performance across different subgroups. For example, patients can be divided into different subgroups based on clinical characteristics such as age, gender, and tumor stage, and the predictive accuracy of the model can be calculated for each subgroup. This stratified analysis helps to understand the model's applicability to different patient populations, providing more targeted guidance for clinical applications.

[0167] S6.2: Construct a comparison model and calculate performance

[0168] To demonstrate the superiority of deep learning models, this application constructs a traditional Cox regression model and a random forest survival model as a comparison. These models use the same feature set and training data, and their C-index values ​​are calculated on an independent validation set.

[0169] S6.3: Compare the performance metrics of different models

[0170] By comparing the performance metrics of different models, the advantages of deep learning models can be intuitively demonstrated. Practice shows that deep learning models can usually achieve higher C-index and AUC values, thanks to their ability to capture complex nonlinear relationships in gene expression data.

[0171] S7: Perform external verification

[0172] Specifically, external verification includes:

[0173] S7.1: Obtain multiple external validation sets

[0174] External validation of a model is a crucial step in evaluating its practical application value. To ensure the model has good generalization ability, this application uses multiple independent external validation sets for comprehensive evaluation. These validation sets include data from different ethnic groups, regions, and sequencing platforms, which can fully verify the model's applicability and stability.

[0175] For example, firstly, multiple external validation sets are obtained from the GEO and ICGC databases. These datasets include pancreatic cancer datasets from Asian populations, pancreatic cancer datasets from European and American populations, and data from different sequencing platforms (such as microarray platforms and high-throughput sequencing platforms). Then, each external validation set undergoes the same data preprocessing procedures as the training set, including missing value handling, batch effect elimination, and data normalization. This ensures that the external validation data are comparable to the training data.

[0176] S7.2: Evaluate the model's predictive performance on different data.

[0177] When evaluating a model's predictive performance on different external validation sets, the consistency and stability of model performance are key considerations. Calculating the C-index, time-dependent ROC curves, and survival analysis results for each validation set provides a comprehensive understanding of the model's performance across various datasets. If a significant performance drop is observed on certain validation sets, the cause needs to be analyzed and targeted optimizations implemented. For example, the heterogeneity of the dataset might need to be considered, or the model's parameter settings might need to be adjusted.

[0178] S7.3: Retrain based on external prediction results

[0179] Based on the results of external validation, the model may need to be retrained or fine-tuned. For example, if the model's performance on Asian population datasets is found to be unsatisfactory, it's advisable to include some Asian population data in the training set for model updates. During retraining, it's crucial to maintain the model's generalization ability and avoid overfitting to data from a specific population or platform. Through this iterative optimization process, a model with good predictive capabilities across population groups and platforms can ultimately be obtained.

[0180] External validation results of the model not only reflect its predictive performance but also provide important references for clinical applications. Detailed documentation of the model's performance on different validation sets can help clinicians understand the model's applicability in their specific populations, thereby better guiding individualized treatment decisions.

[0181] S8: Generate Forecast Report

[0182] like Figure 4 The goal of generating personalized prognostic prediction reports is to provide intuitive and interpretable prediction results, offering valuable reference information for clinical decision-making.

[0183] Specifically, generating the forecast report includes:

[0184] S8.1: Calculate gene contribution

[0185] This can be achieved by analyzing the weight parameters and activation values ​​in the deep learning model. Specifically, the Integrated Gradients method is used to calculate the importance score of each input feature to the final prediction result. This method can quantitatively assess the role of each gene in the prediction process, helping to understand the model's decision-making basis. For genes with large contributions, special attention should be paid to their expression in patient samples, as these genes are often key factors affecting prognosis.

[0186] S8.2: Pathway enrichment analysis of high-risk genes

[0187] For high-risk genes whose contribution exceeds a preset threshold, pathway enrichment analysis is performed. This step uses multiple bioinformatics databases, such as KEGG, Reactome, and Gene Ontology, to analyze the biological pathways and functions involved by these genes. The enrichment significance of each pathway is calculated using hypergeometric tests to screen for significantly enriched biological processes. These enrichment analysis results help explain the molecular mechanisms of patient prognostic risk and provide a theoretical basis for the development of subsequent treatment plans.

[0188] S8.3: Generate personalized prognostic prediction reports

[0189] Based on the above analysis results, a comprehensive personalized prognostic prediction report is generated. This report includes the following core components: the patient's predicted survival risk score, survival curve prediction graph, expression level analysis of key prognostic-related genes, enrichment results of important biological pathways, and personalized treatment recommendations based on these findings. The report is presented in a combined text and graphic format, ensuring both professionalism and ease of understanding and use by clinicians. For example, heatmaps can be used to display the expression patterns of key genes, bar charts can be used to show pathway enrichment results, and survival curve graphs can be used to display prognostic prediction results.

[0190] This personalized prognostic report not only helps doctors better understand the prognostic risks of each patient, but also provides data support for developing personalized treatment plans. This precision medicine approach based on multi-omics data holds the promise of improving treatment outcomes and quality of life for pancreatic cancer patients.

[0191] Example 2

[0192] like Figure 5 As shown in the embodiments of this application, an apparatus for constructing a prognostic model of pancreatic cancer ferroptosis-related genes is also provided, comprising:

[0193] Data acquisition module 10 is used to acquire gene expression datasets and clinical information datasets of pancreatic cancer patients;

[0194] The preprocessing module 20 is used to perform data preprocessing on the gene expression dataset and the clinical information dataset to generate a standardized gene expression matrix;

[0195] The feature selection module 30 is used to perform differential expression analysis and survival analysis based on the standardized gene expression matrix and the ferroptosis pathway gene set, and to screen out the ferroptosis feature gene set that is significantly related to prognosis.

[0196] Model building module 40 is used to build a deep learning prognostic prediction model based on the ferroptosis feature gene set;

[0197] The model evaluation module 50 is used to evaluate the performance of the deep learning prognostic prediction model using an independent validation set and obtain model evaluation metrics.

[0198] This application embodiment also provides a computer device, the computer device comprising:

[0199] At least one processor; and,

[0200] A memory communicatively connected to the at least one processor; wherein,

[0201] The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform the above-described method for constructing a pancreatic cancer ferroptosis-related gene prognostic model based on gene expression data.

[0202] This application also provides a computer-readable storage medium storing computer instructions for causing a computer to execute the above-described method for constructing a pancreatic cancer ferroptosis-related gene prognostic model based on gene expression data.

[0203] This application also provides a computer program product, including computer instructions, characterized in that, when executed by a processor, the computer instructions implement the steps of the above-described method for constructing a prognostic model of pancreatic cancer ferroptosis-related genes based on gene expression data.

[0204] The terms "first," "second," "third," "fourth," "1," "2," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in a sequence other than that shown in the figures or text.

[0205] It should be understood that although arrows indicate various operation steps in the flowcharts of this application's embodiments, the order in which these steps are implemented is not limited to the order indicated by the arrows. Unless explicitly stated herein, in some implementation scenarios of this application's embodiments, the implementation steps in each flowchart can be executed in other orders as required. Furthermore, some or all steps in each flowchart, based on the actual implementation scenario, may include multiple sub-steps or multiple stages. Some or all of these sub-steps or stages can be executed at the same time, and each sub-step or stage can also be executed at different times. In scenarios where execution times differ, the execution order of these sub-steps or stages can be flexibly configured according to requirements, and this application's embodiments do not limit this.

[0206] The above description is only an optional implementation method for some implementation scenarios of this application. It should be noted that for those skilled in the art, other similar implementation methods based on the technical concept of this application without departing from the technical concept of this application also fall within the protection scope of the embodiments of this application.

Claims

1. A method for constructing a prognostic model of pancreatic cancer ferroptosis-related genes, characterized in that, include: Obtain gene expression datasets and clinical information datasets from pancreatic cancer patients; Data preprocessing is performed on the gene expression dataset and the clinical information dataset to generate a standardized gene expression matrix; Based on the standardized gene expression matrix, combined with the ferroptosis pathway gene set, differential expression analysis and survival analysis were performed to screen out the ferroptosis characteristic gene set that was significantly associated with prognosis. A deep learning prognostic prediction model was constructed based on the aforementioned ferroptosis feature gene set. The performance of the deep learning prognostic prediction model was evaluated using an independent validation set to obtain model evaluation metrics; The differential expression analysis, based on the standardized gene expression matrix and the ferroptosis pathway gene set, includes: The standardized gene expression matrix is ​​matched with the ferroptosis pathway gene set; The expression difference of each matched gene between tumor tissue and normal tissue was calculated to obtain the differential expression analysis results. The FDR values ​​of the differential expression analysis results were calculated using the Benjamini-Hochberg method; The differential expression analysis results were corrected based on the FDR values; The survival analysis, based on the standardized gene expression matrix and the ferroptosis pathway gene set, includes: Univariate Cox regression analysis was performed on each matched gene to obtain the Cox regression analysis results; Based on the Cox regression analysis results, genes that are significantly associated with survival time were screened out; Kaplan-Meier survival analysis was performed on genes that were significantly associated with survival time to obtain the survival analysis results; The deep learning prognostic prediction model constructed based on the ferroptosis feature gene set includes: Design a deep learning model architecture, which includes an input layer, a fully connected layer, a Dropout layer, and an output layer. The dimension of the input layer is the same as the number of genes in the ferroptosis feature gene set. By combining the Cox partial likelihood loss function and the ranking loss function, a multi-task learning objective for the deep learning model architecture is defined. Add L1 and L2 regularization terms to the multi-task learning objective; Based on the multi-task learning objective and the mini-batch stochastic gradient descent algorithm, the deep learning model architecture is trained and optimized to obtain a deep learning prognostic prediction model.

2. The method according to claim 1, characterized in that, Data preprocessing of the gene expression dataset includes: Remove gene expression data from the gene expression dataset where the proportion of missing values ​​exceeds 20%; The missing values ​​of the remaining gene expression data were filled using the KNN algorithm; The ComBat algorithm was used to standardize the filled gene expression data. Data preprocessing of the clinical information dataset includes: The unit for calculating survival time in the clinical information data of the aforementioned clinical information dataset is unified to months; Encode the survival status in the clinical information data; Integrate the patient's clinical characteristics information from the aforementioned clinical information data; Generate a standardized gene expression matrix, including: The processed gene expression data and clinical information data are matched according to the sample ID to generate a standardized gene expression matrix.

3. The method according to claim 1, characterized in that, The set of ferroptosis-related genes that were screened and were significantly associated with prognosis includes: Based on the results of differential expression analysis and survival analysis, the basic framework of the LASSO regression model was determined. Set the candidate range for the penalty parameter λ of the basic framework; For the ferroptosis pathway gene set, the optimal λ value was selected using a 10-fold cross-validation method; The LASSO regression model was retrained on the ferroptosis pathway gene set using the optimal λ value to obtain the regression coefficient for each gene. Features of genes with non-zero regression coefficients are extracted to form a set of ferroptosis characteristic genes.

4. The method according to claim 1, characterized in that, The evaluation of the performance of the deep learning prognostic prediction model using an independent validation set includes: Calculate the model performance metrics of the deep learning prognostic prediction model on an independent validation set, wherein the model performance metrics include: C-index, time-dependent ROC curve, and survival analysis; Construct Cox regression and random forest survival models using data from independent validation sets, respectively. Calculate the C-index of the Cox regression model on the independent validation set, and the C-index of the random forest survival model on the independent validation set; Compare the C-index of the deep learning prognostic prediction model, the C-index of the Cox regression model, and the C-index of the random forest survival model; The performance of the deep learning prognostic prediction model was determined based on the comparison results, time-dependent ROC curves, and survival analysis.

5. The method according to claim 4, characterized in that, Also includes: Multiple external validation sets are obtained, including external validation sets of data from different ethnic groups and external validation sets of data from different sequencing platforms. The external prediction performance of the deep learning prognostic prediction model on data from different ethnic groups and different sequencing platforms was evaluated based on multiple external validation sets. Based on the external prediction results, the deep learning prognostic prediction model is retrained.

6. The method according to claim 1, characterized in that, Also includes: Calculate the contribution of each gene to the prediction results; Pathway enrichment analysis was performed on high-risk genes whose contribution exceeded the threshold. Based on the pathway enrichment analysis results, a personalized prognostic prediction report is generated.

7. A device for constructing a prognostic model of pancreatic cancer ferroptosis-related genes, characterized in that, include: The data acquisition module is used to acquire gene expression datasets and clinical information datasets from pancreatic cancer patients. The preprocessing module is used to perform data preprocessing on the gene expression dataset and the clinical information dataset to generate a standardized gene expression matrix; The feature selection module is used to perform differential expression analysis and survival analysis based on the standardized gene expression matrix and the ferroptosis pathway gene set, and to screen out the ferroptosis feature gene set that is significantly related to prognosis. The model building module is used to build a deep learning prognostic prediction model based on the ferroptosis feature gene set. The model evaluation module is used to evaluate the performance of the deep learning prognostic prediction model using an independent validation set and obtain model evaluation metrics. The differential expression analysis, based on the standardized gene expression matrix and the ferroptosis pathway gene set, includes: The standardized gene expression matrix is ​​matched with the ferroptosis pathway gene set; The expression difference of each matched gene between tumor tissue and normal tissue was calculated to obtain the differential expression analysis results. The FDR values ​​of the differential expression analysis results were calculated using the Benjamini-Hochberg method; The differential expression analysis results were corrected based on the FDR values; The survival analysis, based on the standardized gene expression matrix and the ferroptosis pathway gene set, includes: Univariate Cox regression analysis was performed on each matched gene to obtain the Cox regression analysis results; Based on the Cox regression analysis results, genes that are significantly associated with survival time were screened out; Kaplan-Meier survival analysis was performed on genes that were significantly associated with survival time to obtain the survival analysis results; The deep learning prognostic prediction model constructed based on the ferroptosis feature gene set includes: Design a deep learning model architecture, which includes an input layer, a fully connected layer, a Dropout layer, and an output layer. The dimension of the input layer is the same as the number of genes in the ferroptosis feature gene set. By combining the Cox partial likelihood loss function and the ranking loss function, a multi-task learning objective for the deep learning model architecture is defined. Add L1 and L2 regularization terms to the multi-task learning objective; Based on the multi-task learning objective and the mini-batch stochastic gradient descent algorithm, the deep learning model architecture is trained and optimized to obtain a deep learning prognostic prediction model.

Images