Method and system for constructing an acute leukemia risk prediction model based on transcriptome data

Abstract

The application discloses a kind of based on transcriptome data's acute leukemia risk prediction model construction method and system, method includes the following steps: obtaining transcriptome sequencing original count matrix and sample label;Original count matrix is executed washing, normalization and logarithmic conversion, generate standard training expression characteristic matrix;Standard training expression characteristic matrix is carried out batch effect correction;Based on the training expression characteristic matrix after correction, using multi-stage feature selection strategy, screen out the core feature gene set related to acute leukemia discrimination;With the expression amount of core feature gene set as input feature, with sample label as prediction target, using LASSO regularized logistic regression model is trained to obtain acute leukemia risk probability prediction model;Risk probability prediction model is evaluated using test set to evaluate the discriminant ability and generalization performance of the model.The method can improve the accuracy, stability and practicability of acute leukemia diagnosis.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and medical artificial intelligence, and in particular to a method and system for constructing an acute leukemia risk prediction model based on transcriptome data. Background Technology

[0002] Acute leukemia is a group of malignant hematologic disorders characterized by rapid onset, rapid progression, and high mortality. It mainly includes acute myeloid leukemia (AML) and acute lymphoblastic leukemia (ALL). These diseases result from the malignant clonal proliferation of hematopoietic stem cells or progenitor cells, severely suppressing normal hematopoietic function. Patients can develop severe clinical symptoms such as anemia, infection, and bleeding within a short period. If timely and accurate diagnosis and targeted treatment are not provided, the prognosis is often poor.

[0003] In clinical practice, there are significant differences in treatment plans, prognostic assessments, and medication strategies between AML and ALL. Therefore, rapid and accurate subtyping diagnosis of acute leukemia is a crucial prerequisite for clinical decision-making. Currently, routine diagnostic methods for acute leukemia mainly include peripheral blood or bone marrow smear morphological examination, immunophenotypic analysis, cytogenetic testing, and molecular biological testing. However, these methods still have many shortcomings in practical application.

[0004] First, morphological examination relies on experienced hematologists who interpret cell morphology under a microscope, making it highly subjective and with limited accuracy in diagnosing early or atypical cases. Second, immunophenotyping analysis typically depends on flow cytometry, requiring multiple antibody combinations, which is not only costly but also suffers from significant differences in testing platforms and antibody systems between different laboratories, resulting in insufficient reproducibility and standardization of results. While cytogenetic and molecular biological testing are valuable in typing and prognostic assessment, their procedures are complex and time-consuming, and they often target a limited number of known mutation sites, making it difficult to comprehensively reflect the molecular characteristics of the disease.

[0005] With the development of high-throughput sequencing technology, RNA sequencing (RNA-seq) technology can systematically and comprehensively reflect gene expression status at the transcriptome level, providing a new technical means for molecular subtyping and mechanism research of leukemia. Numerous studies have shown significant differences in gene expression profiles between AML and ALL, and these differences contain important diagnostic information. However, RNA-seq data typically has characteristics such as high dimensionality, relatively limited sample size, and significant batch effects. Directly using it for modeling analysis can easily lead to overfitting problems and makes it difficult to maintain stable performance across different data sources and detection platforms.

[0006] In recent years, machine learning and artificial intelligence technologies have received widespread attention in the field of medical data analysis, particularly demonstrating unique advantages in feature selection and classification modeling of high-dimensional omics data. For example, regularization methods such as LASSO (Least Absolute Shrinkage and Selection Operator) can achieve feature selection during model training, effectively reducing model complexity and improving model generalization ability. However, existing machine learning-based molecular diagnostic research on leukemia often focuses on single disease types or relies on specific datasets for modeling, generally suffering from insufficient batch effect handling, inadequate model stability validation, and limited clinical generalizability. Furthermore, in actual clinical applications, patient samples often come from different sequencing platforms, different experimental batches, or different populations, inevitably leading to systematic biases in the data. Without standardized data preprocessing procedures and robust model building strategies, the diagnostic performance of the model on independent datasets can easily decline significantly, thus limiting its clinical application value.

[0007] Therefore, there is an urgent need for a molecular diagnostic method and system for acute leukemia that combines bioinformatics and medical artificial intelligence technologies. This method should be able to accurately differentiate and diagnose acute leukemia, especially AML and ALL, based on RNA-seq data, through standardized data preprocessing, effective batch effect correction, and robust machine learning modeling strategies. The method should possess good reproducibility, cross-platform stability, and clinical scalability to meet the needs of practical clinical auxiliary diagnosis. Summary of the Invention

[0008] This invention aims to address at least one of the technical problems existing in the prior art. To this end, this invention proposes a method for constructing an acute leukemia risk prediction model based on transcriptome data. This method achieves automated and intelligent molecular diagnosis of acute leukemia samples by constructing a standardized transcriptome data processing workflow and introducing a regularized feature selection and classification model, thereby overcoming the shortcomings of existing technologies in terms of accuracy, stability, and practicality.

[0009] This invention also proposes a system for constructing an acute leukemia risk prediction model based on transcriptome data.

[0010] The method for constructing an acute leukemia risk prediction model based on transcriptome data according to the first aspect of the present invention includes the following steps performed by a computer: S1. Training data acquisition and construction steps: Obtain raw counting matrices of peripheral blood transcriptome sequencing from multiple sources and corresponding sample labels, including acute leukemia sample labels and non-acute leukemia sample labels; S2. Data preprocessing steps: Cleaning, normalization, and logarithmic transformation are performed on the original counting matrix to generate a standardized training representation feature matrix; S3. Batch effect correction step: Perform batch effect detection and correction on the standardized training expression feature matrix to eliminate systematic bias between different data sources or sequencing platforms, while retaining biological difference information related to acute leukemia. S4. Feature selection step: Based on the corrected training expression feature matrix, a multi-stage feature selection strategy is adopted to select the core feature gene set related to acute leukemia discrimination; S5. Model training steps: Using the expression level of the core feature gene set as the input feature and the sample label as the prediction target, train the acute leukemia risk probability prediction model using the LASSO regularized logistic regression model. S6. Model evaluation steps: Use the test set to evaluate the performance of the trained risk probability prediction model, and calculate the area under the receiver operating characteristic curve, accuracy, sensitivity and specificity index to evaluate the model's discriminative ability and generalization performance.

[0011] In some embodiments, step S2 specifically includes: Remove genes whose total expression is zero or whose expression variance is zero in all samples; The TMM method was used to normalize the sample size factor. The normalized expression level is converted into logarithmic CPM form.

[0012] In some embodiments, the batch effect correction in step S3 employs the empirical Bayesian ComBat model, whose generative model is expressed as follows: in, For the first The gene in the first Standardized expression values ​​in each sample, Indicates the first The overall average expression level of each gene across all samples; Indicates the first The data batch to which each sample belongs; Indicates batch On genes The introduced additive shift effect; Indicates batch On genes The introduced multiplicative scaling effect; This is a random error term with a mean of zero.

[0013] In some embodiments, the multi-stage feature selection strategy in step S4 includes: First screening stage: Use the random forest classification model to calculate the Gini importance of each gene, sort them in descending order of importance, and screen out the top N genes to form a preliminary feature gene set; The second screening stage: On the preliminary feature gene set, a LASSO-regularized logistic regression model is used for training. The optimal regularization parameter is selected through 10-fold cross-validation, and genes with non-zero regression coefficients are identified as the core feature gene set.

[0014] In some embodiments, the loss function of the LASSO regularized logistic regression model is: in: It is the first The true label of each sample; It is through the current model to the first Predicted values ​​for each sample; It is a regularization parameter that controls the intensity of feature selection; It is the first The regression coefficient of each gene; It is the total number of features; It represents the total number of samples.

[0015] In some embodiments, the core feature gene set includes multiple genes from the following: NF1, SF3B1, CUX1, CDKN2A, HOXA9, SPI1, MYD88, FLT3, STAT5B, ERG, PIK3CA, IDH1, WT1, MN1, RARA, CXCR4, DDX41, BCOR, HOXA10, DUX4, MCL1, CDKN2B, JAK1, CSF3R.

[0016] In some embodiments, step S6 is followed by: Model saving steps: Save the model parameters of the trained risk probability prediction model, the list of core feature genes, and the performance evaluation results in a unified manner to ensure the traceability and reproducibility of the model.

[0017] In some embodiments, the training data in step S1 comes from multiple independent data sources, including at least the TARGET, TCGA-LAML, GTEx, and GSE227832 queues; the acute leukemia samples include acute myeloid leukemia samples and acute lymphoblastic leukemia samples.

[0018] The acute leukemia risk prediction model construction system based on transcriptome data according to the second aspect of the present invention includes: The data acquisition module is used to acquire raw counting matrices of peripheral blood transcriptome sequencing from multiple sources and their corresponding sample labels; The data preprocessing module is used to perform cleaning, normalization, logarithmic transformation and batch effect correction on the original counting matrix to generate a standardized training representation feature matrix. The feature selection module is used to select the core feature gene set based on the corrected training expression feature matrix, using a multi-stage strategy that combines random forest and LASSO regularized logistic regression. The model training module is used to train an acute leukemia risk probability prediction model by taking the expression level of the core feature gene set as input, the sample label as the prediction target, and using a LASSO-regularized logistic regression algorithm and a cross-validation strategy. The model evaluation module is used to evaluate the performance of the trained risk probability prediction model using a test set and output evaluation metrics.

[0019] In some embodiments, the system is encapsulated in a server, cloud computing platform, or medical artificial intelligence inference terminal, and outputs the risk probability prediction model and its evaluation results through a graphical interface or application programming interface.

[0020] The present invention has at least the following beneficial effects: 1. By employing data preprocessing and batch effect correction steps, we ensure that training and testing data are consistent in gene space and expression scale, thereby significantly improving the model's cross-cohort generalization ability.

[0021] 2. By employing a multi-stage feature selection strategy, namely a two-stage screening using random forest and LASSO, this invention automatically identifies and robustly selects a small subset of genes most relevant to acute leukemia from a high-dimensional expression matrix, thereby improving model interpretability and clinical usability.

[0022] 3. Save the best model, the list of selected genes, and the complete validation results file to ensure the reproducibility of the model results and the traceability of the review.

[0023] 4. The method of the present invention is applicable to peripheral blood RNA-seq data sequenced on different age groups (children, adults, and the elderly) and different platforms, which facilitates its application in multi-center and multi-platform environments.

[0024] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating the steps of a method for constructing an acute leukemia risk prediction model based on transcriptome data according to an embodiment of the present invention. Figure 2 This is a schematic flowchart of a method for constructing an acute leukemia risk prediction model based on transcriptome data according to an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the calculation of predicted probability output using a training set and a test set according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the AUC performance output of the 44th iteration according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the ROC curve of the 44th iteration training set according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the ROC curve of the 44th iteration test set according to an embodiment of the present invention; Figure 7 This is the gene risk forest diagram of the 44th iteration according to an embodiment of the present invention; Figure 8 This is a principal component analysis (PCA) scatter plot of the expression feature matrix before batch effect correction according to an embodiment of the present invention, colored by data batch; Figure 9 This is a principal component analysis scatter plot of the expression feature matrix before batch effect correction according to an embodiment of the present invention, colored by disease state (acute leukemia and non-acute leukemia); Figure 10 This is a principal component analysis scatter plot of the batch-effect corrected expression feature matrix according to a batch of data, based on an embodiment of the present invention. Figure 11 This is a principal component analysis scatter plot of the batch effect-corrected expression feature matrix colored by disease state according to an embodiment of the present invention. Detailed Implementation

[0027] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0028] The following disclosure provides numerous different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples. Such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. Additionally, examples of various specific processes and materials are provided in this invention; however, those skilled in the art will recognize the applicability of other processes and / or the use of other materials.

[0029] Below, in conjunction with Figure 1 and Figure 2 According to the method for constructing an acute leukemia risk prediction model based on transcriptome data according to the first aspect of the present invention, the method includes the following steps performed by a computer: S1. Training data acquisition and construction steps: Obtain raw counting matrices of peripheral blood transcriptome sequencing from multiple sources and corresponding sample labels, including acute leukemia sample labels and non-acute leukemia sample labels.

[0030] Raw RNA-seq counting matrices were collected from peripheral whole blood samples from multiple sources. Specifically, peripheral blood samples were collected following clinical procedures, and counting matrices were obtained by following the steps of blood collection, RNA extraction, library construction, and high-throughput sequencing. Training data were obtained from multiple public databases (such as TCGA and GTEx) or clinical cohorts and were confirmed by experts as AML, ALL, or healthy labels.

[0031] Furthermore, data from different sources are unified into a standardized gene × sample matrix for subsequent feature processing and model training.

[0032] in Indicates the number of genes. Indicates the number of samples. Indicates the first The gene in the first Raw sequencing count values ​​from each sample.

[0033] Each sample corresponds to one disease label: in, This indicates an acute leukemia sample. This indicates a peripheral blood sample from a non-acute leukemia patient.

[0034] S2. Data preprocessing steps: Clean, normalize, and logarithmically transform the original counting matrix to generate a standardized training representation feature matrix.

[0035] Specifically, step S2 includes: Remove genes whose total expression is zero or whose expression variance is zero in all samples; The TMM method was used to normalize the sample size factor. The normalized expression level is converted into logarithmic CPM form.

[0036] For example, the data preprocessing steps include removing low-information features from the original counting matrix, deleting genes whose total expression or expression variance is zero across all samples; based on this, the TMM method is used to normalize the sample size factor. Specifically, using a reference sample as a benchmark, the log expression ratio (M value) and average expression intensity (A value) of each gene between the sample to be normalized and the reference sample are calculated, and extreme M values ​​and A values ​​are pruned; based on this, the M values ​​of the remaining genes are weighted and averaged to estimate the relative library scaling factor of the sample, thereby correcting for systematic biases caused by differences in sequencing depth and RNA composition between different samples.

[0037] Furthermore, the normalized expression level was converted into log CPM (logCPM) form to eliminate systematic bias caused by differences in sequencing depth and RNA composition.

[0038] S3. Batch effect correction step: Batch effect detection and correction are performed on the standardized training expression feature matrix to eliminate systematic biases between different data sources or sequencing platforms, while retaining biological differences related to acute leukemia.

[0039] For example, principal component analysis is performed on the expression matrix after merging multiple cohorts to assess batch effects; after confirming the existence of batch effects, the ComBat batch correction model is used to correct for systematic biases introduced by different sequencing platforms or data sources. Technical batch differences are eliminated to obtain a corrected expression matrix with a uniform scale. Specifically, for the first... The gene in the first The standardized expression value in each sample is denoted as . Its generative model is represented as: in: Indicates the first The overall average expression level of each gene across all samples; Indicates the first The data batch to which each sample belongs; Indicates batch On genes The introduced additive shift effect; Indicates batch On genes The introduced multiplicative scaling effect; This is a random error term with a mean of zero.

[0040] Furthermore, to improve the stability of parameter estimation under conditions of multiple batches, small samples, or high noise, an empirical Bayesian method is used to estimate the batch effect parameters. Specifically, it is assumed that the additive shift parameter and multiplicative scaling parameter of different genes in the same batch follow the following prior distributions: By sharing prior information among all genes, the batch effect parameters of each gene in different batches are estimated in a contracted manner, thereby obtaining robust parameter estimation results.

[0041] After completing the parameter estimation above, batch correction is performed on the original standardized expression values ​​to obtain the corrected expression matrix, whose expression values ​​are calculated as follows: in This represents the batch-corrected gene expression values; Furthermore, the training set is selected as the reference data batch and the standard batch, and its corresponding additive and multiplicative batch effect parameters are fixed to zero and one, respectively, so that the expression distribution of the remaining batches is uniformly mapped to the scale space of the reference batch.

[0042] S4. Feature selection step: Based on the corrected training expression feature matrix, a multi-stage feature selection strategy is adopted to select the core feature gene set related to acute leukemia discrimination.

[0043] Specifically, the stage feature selection strategies include: First screening stage: Use the random forest classification model to calculate the Gini importance of each gene, sort them in descending order of importance, and screen out the top N genes to form a preliminary feature gene set; The second screening stage: On the initial feature gene set, a LASSO-regularized logistic regression model is used for training. The optimal regularization parameter is selected through 10-fold cross-validation, and genes with non-zero regression coefficients are identified as the core feature gene set.

[0044] For example, a random forest classification model is constructed using training set data, with acute leukemia and non-disease status as classification labels, and the Gini importance of each gene in the model is calculated; the genes with the highest ranking are selected according to importance to form a preliminary feature gene set.

[0045] Based on the aforementioned preliminary feature gene set, a LASSO-regularized logistic regression model is introduced. By applying L1 sparsity constraints to the regression coefficients and combining them with cross-validation strategies to optimize the regularization parameters, genes with non-zero regression coefficients are automatically selected, thus constructing a core feature gene set with high stability and high discriminative power.

[0046] Specifically, the logistic regression model based on LASSO regularization further filters features to improve the model's interpretability and accuracy. The LASSO loss function minimization formula is as follows: in: It is the first The true label of each sample; It is through the current model to the first Predicted values ​​for each sample; It is a regularization parameter that controls the intensity of feature selection; It is the first The regression coefficient of each gene; It is the total number of features (in this example, the number of genes in the SRF). It represents the total number of samples.

[0047] Through the aforementioned random forest importance screening and LASSO regularized logistic regression joint feature selection process, the core feature gene set consists of genes with non-zero regression coefficients in the final LASSO model. It exhibits stable feature selection consistency and significant discriminative ability under multiple random resampling and multi-cohort data conditions, and the relative contribution of each gene can be visualized using a forest graph. The core feature gene set includes the following 24 genes: NF1, SF3B1, CUX1, CDKN2A, HOXA9, SPI1, MYD88, FLT3, STAT5B, ERG, PIK3CA, IDH1, WT1, MN1, RARA, CXCR4, DDX41, BCOR, HOXA10, DUX4, MCL1, CDKN2B, JAK1, and CSF3R. The gene risk forest graph constructed based on the core feature genes is shown below. Figure 7 As shown in the figure, this figure is used to illustrate the risk contribution direction and relative magnitude of each core feature gene in the acute leukemia probability prediction model.

[0048] Specifically, CUX1 is a transcription factor involved in cell cycle regulation, DNA damage repair, and the maintenance of hematopoietic homeostasis. Previous studies have shown that CUX1 plays an important regulatory role in the balance of proliferation and differentiation of hematopoietic stem cells and progenitor cells during normal hematopoiesis, and its weakened function has been shown to promote abnormal proliferation of hematopoietic progenitor cells and increase the risk of leukemia. Therefore, CUX1 is considered a dose-sensitive tumor suppressor gene in acute myeloid leukemia. NF1 encodes neurofibromatosis protein and is a key negative regulator of the Ras signaling pathway, participating in the regulation of hematopoietic stem and progenitor cell proliferation and differentiation during normal hematopoiesis. Studies have shown that loss of NF1 function or downregulation of expression can lead to excessive activation of Ras signaling, promoting abnormal proliferation of hematopoietic progenitor cells and significantly increasing the risk of acute myeloid leukemia. Therefore, NF1 is considered an important tumor suppressor gene in acute leukemia. SF3B1 is a core component of the spliceosome, participating in the splicing of precursor mRNA and playing a crucial role in maintaining the accuracy of gene expression in hematopoietic cells. Numerous studies have confirmed the high frequency of SF3B1 mutations in myelodysplastic syndromes and acute leukemia. These abnormalities can lead to missplicing of key hematopoietic regulatory genes, thereby promoting leukemia development. SF3B1 is now considered a splicing factor of significant clinical importance in myeloid tumors. IDH1 encodes a key enzyme involved in the regulation of cellular metabolism, playing a vital role in cellular metabolic homeostasis and redox balance under normal conditions. Studies have shown that IDH1 mutations or abnormal expression can serve as important criteria for molecular subtyping and risk assessment of leukemia. Especially in patients with acute myeloid leukemia (AML), IDH1 abnormalities have been incorporated into clinical molecular testing systems to aid in disease classification and prognosis. IDH1 is considered a highly relevant molecular biomarker in AML. HOXA9 is an important member of the homeobox transcription factor family, playing a crucial role in the regulation of hematopoietic stem cell self-renewal and differentiation. Studies have shown that HOXA9 is highly expressed in various acute leukemia subtypes, especially in patients with mixed lineage leukemia rearrangements. Its aberrant activation can maintain the undifferentiated state of leukemia cells and enhance their proliferative capacity. HOXA9 is considered a core transcriptional regulator in the occurrence and maintenance of acute leukemia. CDKN2A is a classic cell cycle regulatory gene and a target of the PML / RARα tumor fusion protein in the development of acute promyelocytic leukemia. It plays a central role in cell cycle checkpoint control and cell proliferation inhibition. Studies have shown that CDKN2A is frequently lost, inactivated, or underexpressed in acute lymphoblastic leukemia, and its functional defects can lead to abnormal proliferation of hematopoietic cells and promote leukemia development. FLT3 is a gene encoding a receptor-type tyrosine kinase that participates in the regulation of hematopoietic stem and progenitor cell survival, proliferation, and differentiation during normal hematopoiesis. Studies have shown that FLT3 abnormalities have a high incidence in AML patients and have been widely used in molecular subtyping, risk stratification, and treatment decisions for AML. FLT3 is one of the most common and clinically significant molecular markers in acute myeloid leukemia.

[0049] In this invention, the expression characteristics of genes such as NF1, SF3B1, CUX1, CDKN2A, HOXA9, SPI1, and FLT3 showed stable differences between diseased and non-disease-related peripheral blood samples, indicating their good disease discrimination value at the peripheral blood transcriptome level. The core molecular diagnostic genes identified above are all involved in hematopoietic differentiation, leukemia development, or leukemia-related signaling pathways, and their abnormal expression in peripheral blood can accurately reflect changes in leukemia cell burden or disease state. By incorporating genes with clear leukemia-related characteristics into the model's input, the interpretability and clinical acceptability of the model are significantly enhanced while improving discriminative performance.

[0050] S5. Model training steps: Using the expression level of the core feature gene set as the input feature and the sample label as the prediction target, the model is trained using a LASSO-regularized logistic regression model to obtain the acute leukemia risk probability prediction model. First, the training and test sets are divided in an 8:2 ratio to ensure sufficient samples in each batch and disease group. The training set includes samples from different batches and clinical backgrounds to guarantee the model's broad applicability.

[0051] The logistic regression model is trained using LASSO regularization, and its formula is as follows: in: It is a sample Predicted probability of belonging to the target category (illness or health); It is the input feature (i.e., the expression level of the selected genes). These are regression coefficients (which are optimized using a LASSO regularized logistic regression model). The training rounds are set to 100. During each training run, the optimal hyperparameters (including the regularization coefficient λ) are determined through cross-validation. This step ensures the model's generalization ability and stability.

[0052] Therefore, this step uses the expression levels of the core feature gene set as input features and the sample labels as prediction targets to train an acute leukemia risk probability prediction model. The final LASSO logistic regression probability predictor model is then trained using the selected core feature gene set. Based on the predicted probability values, classification is performed using the following rules: if the predicted probability... If the value is greater than the set threshold of 0.5, the sample is determined to be either AML or ALL; if the predicted probability... If the value is less than 0.5, it is considered a non-disease.

[0053] S6. Model Evaluation Steps: Use the test set to evaluate the performance of the trained risk probability prediction model, and calculate the area under the receiver operating characteristic curve, accuracy, sensitivity, and specificity to evaluate the model's discriminative ability and generalization performance.

[0054] In each iteration, predicted probabilities are calculated for both the training and test sets, and the model's discriminative ability is evaluated based on the receiver operating characteristic (ROC) curve, with the corresponding area under the curve (AUC) calculated. Simultaneously, the predicted probabilities are classified using a fixed threshold, and the model's accuracy in the diseased and non-disease classes, as well as the overall accuracy, are calculated. The model's performance on both the training and test sets is evaluated using metrics such as AUC, accuracy, sensitivity, and specificity to assess its discriminative ability and generalization performance.

[0055] In some embodiments, step S6 is followed by: Model saving steps: Save the model parameters, core feature gene list, and performance evaluation results of the trained risk probability prediction model in a unified manner to ensure the traceability and reproducibility of the model.

[0056] The following describes, with reference to specific embodiments, the method for constructing an acute leukemia risk prediction model based on transcriptome data according to the first aspect of the present invention.

[0057] S1. Steps for acquiring and constructing training data: In this embodiment, the peripheral blood RNA sequencing data used were obtained from multiple public databases and research cohorts, including: Peripheral blood samples from the TARGET project included 386 AML (acute myeloid leukemia) patients and 31 non-AML control samples. A total of 151 peripheral blood or bone marrow-derived AML samples were collected from the TCGA-LAML project. Peripheral blood samples from healthy individuals in the GTEx project, totaling 337, were used as a healthy control group. The GSE227832 dataset contains peripheral blood samples from ALL (acute lymphoblastic leukemia) patients. Two ALL patients in remission were removed from the dataset. The dataset contains 328 ALL patients and 10 non-ALL controls, for a total of 338 samples, which are used to construct control data for ALL patients.

[0058] In this embodiment, all data used are derived from peripheral blood RNA sequencing data that has undergone high-throughput sequencing. This data exists in the raw form of a gene counting matrix, with each column corresponding to an independent subject sample and each row corresponding to a gene transcript. The raw counting matrix is ​​downloaded from a public database, stored in the native R language data format (Rdata), and loaded into the computing environment using R functions.

[0059] S2. Data preprocessing steps: After the data is loaded, low-information features are first removed from each dataset.

[0060] Specifically, for any gene, if the sum of its expression counts in all samples is zero, the gene is considered to have no discriminative information and is deleted. Furthermore, genes with zero expression variance in all samples are also removed to avoid redundant features or numerical instability in subsequent model training.

[0061] After the above filtering was performed, the gene expression matrix was normalized using the edgeR software package using TMM. Furthermore, the normalized counts were converted to log CPM (parts per million).

[0062] S3. Batch effect correction steps: To assess the systematic shifts among multi-cohort data, principal component analysis (PCA) was performed on the processed expression matrix. By visualizing the samples in the principal component space by batch label and disease state, the clustering trends of the samples in terms of technological origin and biological status can be intuitively observed.

[0063] Specifically, after confirming the existence of batch effects, the ComBat algorithm was used to correct for batch effects in the expression matrix, using the TARGET queue as a reference batch to adjust the expression distribution of samples from different sources. After correcting for batch effects using the standard ComBat model, batch-related clustering was significantly reduced.

[0064] After correction, PCA analysis was performed again to verify whether batch-related clusters were significantly reduced while disease-related structures were preserved.

[0065] To evaluate the practical effectiveness of the batch effect correction method described above, this embodiment performs principal component analysis (PCA) on the expression feature matrices before and after correction, and performs two-dimensional visualization according to data batch and disease state. The results are as follows: Figures 8 to 11 As shown.

[0066] Figure 8 and Figure 9 This is a PCA scatter plot before batch effect correction. Among them, Figure 8 Coloring by data batch reveals a clear separation trend among samples from different batches in the principal component space. The first principal component (PC1, explaining 46.6% of the variance) and the second principal component (PC2, explaining 15% of the variance) can clearly distinguish samples from different sequencing platforms or data sources, indicating that there is a significant systematic batch bias in the original data. Figure 9 Coloring by disease state reveals that the acute leukemia samples (AML+ALL) and the non-acute leukemia control samples do not show a clear boundary in the same principal component space. The distributions of the two groups of samples highly overlap, indicating that the batch effect has seriously interfered with the real biological signals related to the disease.

[0067] Figure 10 and Figure 11 This is a PCA scatter plot after ComBat batch effect correction. Figure 10 Coloring by batch shows that the distribution of samples from each batch in the principal component space has become mixed and no longer forms obvious batch clusters. The decrease in variance contribution of the first principal component (PC1, explaining 23% of the variance) and the second principal component (PC2, explaining 6.6% of the variance) also indirectly reflects that the systematic technical bias has been effectively removed. Figure 11 Coloring by disease state reveals a clear separation trend between acute leukemia samples and control samples in the PC1 direction. The distribution centroids of the two groups are significantly different, indicating that the differences in disease-related expression that were originally masked by batch effects are highlighted after correction.

[0068] Based on the results of the four figures above, the batch effect correction strategy adopted in this embodiment effectively eliminates technical batch differences while successfully preserving biological differences related to acute leukemia, providing a high-quality, scale-uniform expression feature matrix for subsequent feature selection and model training.

[0069] S4. Feature selection step: Based on the corrected training expression feature matrix, a multi-stage feature selection strategy is adopted to select the core feature gene set related to acute leukemia discrimination; Random Forest was used to model the training set data. Random Forest uses ensemble learning to classify data using multiple decision trees and evaluate the importance of each gene. During training, each decision tree selects the best splitting feature based on the Gini index, and the gene importance is ultimately evaluated by calculating the Gini importance of all trees. Batch-effects corrected data was used as the input feature matrix, and the sample labels (Leukemia and Normal) were used as the target variable. The model classifies the gene expression of each sample. As the Random Forest classification model iterates through training, the Gini importance of each gene is calculated. Gini importance measures the importance of each gene in the splitting decisions of the model and is calculated as the average importance of each gene across all trees, i.e., a weighted average of its splitting contribution across all trees.

[0070] Based on the Gini importance ranking of each gene, the top N genes are selected. These genes constitute the initial feature set SRF, which will be processed as candidate features in the subsequent sparse feature selection.

[0071] LASSO (Minimum Absolute Shrinkage and Selection Operator) achieves feature selection through L1 regularization (L1norm), which can automatically compress the regression coefficients of some features to zero, thereby selecting the most discriminative features.

[0072] The genes in the SRF were trained using a LASSO-regularized logistic regression model. The most important features were selected by minimizing the loss function. The selected genes from the SRF were used as the input feature matrix, and the labels of the training data were used as the output for LASSO regression training. 10-fold cross-validation was used to select the optimal regularization parameters. Cross-validation divides the data into 10 subsets, training with 9 subsets and testing with 1 subset, and finally evaluates the model's performance by calculating AUC or accuracy. Cross-validation selects the optimal subset. After setting the parameters, the model is trained to form the final LASSO model.

[0073] After the LASSO model is trained, gene features with non-zero regression coefficients are obtained. These genes contribute significantly to the model's predictions and are retained as the final core feature gene set. In the trained LASSO model, genes with regression coefficients not equal to zero are selected to form a highly stable and discriminative core feature gene set for subsequent predictions and model applications. In 100 training iterations, this embodiment uses the maximum AUC on the training set as the selection criterion for the optimal model. The final selected model has a training set AUC of 0.9849 and a test set AUC of 0.9632, indicating that the model possesses excellent discriminative ability and generalization performance. The core feature gene set includes the following 24 genes mentioned above: NF1, SF3B1, CUX1, CDKN2A, HOXA9, SPI1, MYD88, FLT3, STAT5B, ERG, PIK3CA, IDH1, WT1, MN1, RARA, CXCR4, DDX41, BCOR, HOXA10, DUX4, MCL1, CDKN2B, JAK1, and CSF3R.

[0074] according to Figure 7 As can be seen, in the gene risk forest diagram, the vertical axis (Gene) represents the core feature genes included in the acute leukemia diagnostic model after model screening; the horizontal axis (Coefficient, logodds) represents the regression coefficient value of the corresponding gene in the logistic regression model, and its value is the log odds ratio; the dashed line at the zero point of the horizontal axis represents the reference line where the regression coefficient is zero.

[0075] When the regression coefficient of a gene is to the right of zero (coefficient greater than zero), it indicates that, under the condition that other characteristics remain unchanged, the increased expression level of this gene is associated with an increased probability of the sample being diagnosed with acute leukemia. These genes are NF1, SF3B1, CUX1, CDKN2A, HOXA9, SPI1, MYD88, FLT3, STAT5B, and ERG. When it is to the left of zero (coefficient less than zero), it indicates that the expression characteristics of this gene are more inclined towards non-acute leukemia samples, which is reflected in the model as protective or control characteristics. These genes are PIK3CA, IDH1, WT1, MN1, RARA, CXCR4, DDX41, BCOR, HOXA10, DUX4, MCL1, CDKN2B, JAK1, and CSF3R.

[0076] S5. Model training steps: Use the selected core feature gene set to iteratively train the final LASSO logistic regression probability predictor model to obtain the acute leukemia risk probability prediction model.

[0077] Based on the predicted probability values, the following rules are used for classification: If the predicted probability... If the value is greater than the set threshold of 0.5, the sample is determined to be either AML or ALL; if the predicted probability... If the value is less than 0.5, it is considered a non-disease.

[0078] S6. Model Evaluation Steps: In each iteration, predicted probabilities are calculated for both the training and test sets, and the model's discriminative ability is evaluated based on the receiver operating characteristic (ROC) curve, with the corresponding area under the curve (AUC) calculated. Simultaneously, the predicted probabilities are classified using a fixed threshold, and the model's accuracy in the diseased and non-disease categories, as well as the overall accuracy, are calculated. The model's performance on the training and test sets is evaluated using metrics such as AUC, accuracy, sensitivity, and specificity.

[0079] according to Figure 3 As can be seen, in 100 repeated random samplings, the model exhibits stable and high discriminative ability on both the training and test sets. On the training set, the LASSO model has a mean AUC of 0.9780 (range 0.9676–0.9849, 95% CI: 0.9773–0.9787), a mean accuracy of 0.9427 (range 0.9266–0.9558), a mean sensitivity of 0.9596 (range 0.9465–0.9697), and a mean specificity of 0.9042 (range 0.8713–0.9274). On the test set, the mean AUC was 0.9702 (range 0.9400–0.9913, 95% CI: 0.9680–0.9725), the mean accuracy was 0.9315 (range 0.8871–0.9677), the mean sensitivity was 0.9455 (range 0.8960–0.9827), and the mean specificity was 0.8993 (range 0.8133–0.9600). All metrics indicate that the model has good generalization performance. In this embodiment, the model with the highest AUC on the training set was used as the selection criterion for the optimal model. The model from the 44th iteration was ultimately selected as the final model, with a training set AUC of 0.9849 and a test set AUC of 0.9632. Figure 4 This is the AUC performance output for this iteration. Figure 5 and Figure 6 These are the ROC curves for the training and test sets in this iteration, respectively. Figure 7 This is the corresponding gene risk forest diagram.

[0080] In some embodiments, step S6 is followed by: Model saving steps: Save the model parameters, core feature gene list, and performance evaluation results of the trained risk probability prediction model in a unified manner to ensure the traceability and reproducibility of the model.

[0081] Specifically, the performance index matrix, gene selection results, feature frequency statistics table, and optimal model parameters of each training iteration are saved in a unified manner to ensure the complete traceability and reproducibility of the model training and validation process.

[0082] The acute leukemia risk prediction model construction system based on transcriptome data according to the second aspect of the present invention includes: The data acquisition module is used to acquire raw counting matrices of peripheral blood transcriptome sequencing from multiple sources and their corresponding sample labels; The data preprocessing module is used to perform cleaning, normalization, logarithmic transformation and batch effect correction on the original counting matrix to generate a standardized training representation feature matrix. The feature selection module is used to select the core feature gene set based on the corrected training expression feature matrix, using a multi-stage strategy that combines random forest and LASSO regularized logistic regression. The model training module is used to train an acute leukemia risk probability prediction model by taking the expression levels of the core feature gene set as input, the sample labels as prediction targets, and using a LASSO-regularized logistic regression algorithm and a cross-validation strategy. The model evaluation module is used to evaluate the performance of the trained risk probability prediction model using a test set and output evaluation metrics.

[0083] In some embodiments, the system is encapsulated in a server, cloud computing platform, or medical artificial intelligence inference terminal, and outputs the risk probability prediction model and its evaluation results through a graphical interface or application programming interface.

[0084] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0085] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for constructing an acute leukemia risk prediction model based on transcriptome data, characterized in that, This includes the following steps performed by a computer: S1. Training data acquisition and construction steps: Obtain raw counting matrices of peripheral blood transcriptome sequencing from multiple sources and corresponding sample labels, including acute leukemia sample labels and non-acute leukemia sample labels; S2. Data preprocessing steps: Cleaning, normalization, and logarithmic transformation are performed on the original counting matrix to generate a standardized training representation feature matrix; S3. Batch effect correction step: Perform batch effect detection and correction on the standardized training expression feature matrix to eliminate systematic bias between different data sources or sequencing platforms, while retaining biological difference information related to acute leukemia. S4. Feature selection step: Based on the corrected training expression feature matrix, a multi-stage feature selection strategy is adopted to select the core feature gene set related to acute leukemia discrimination; S5. Model training steps: Using the expression level of the core feature gene set as the input feature and the sample label as the prediction target, train the acute leukemia risk probability prediction model using the LASSO regularized logistic regression model. S6. Model evaluation steps: Use the test set to evaluate the performance of the trained risk probability prediction model, and calculate the area under the receiver operating characteristic curve, accuracy, sensitivity and specificity index to evaluate the model's discriminative ability and generalization performance.

2. The method according to claim 1, characterized in that, Step S2 specifically includes: Remove genes whose total expression is zero or whose expression variance is zero in all samples; The TMM method was used to normalize the sample size factor. The normalized expression level is converted into logarithmic CPM form.

3. The method according to claim 1, characterized in that, The batch effect correction in step S3 uses the empirical Bayesian ComBat model, whose generative model is expressed as: in, For the first The gene in the first Standardized expression values ​​in each sample, Indicates the first The overall average expression level of each gene across all samples; Indicates the first The data batch to which each sample belongs; Indicates batch On genes The introduced additive shift effect; Indicates batch On genes The introduced multiplicative scaling effect; This is a random error term with a mean of zero.

4. The method according to claim 1, characterized in that, The multi-stage feature selection strategy in step S4 includes: First screening stage: Use the random forest classification model to calculate the Gini importance of each gene, sort them in descending order of importance, and screen out the top N genes to form a preliminary feature gene set; The second screening stage: On the preliminary feature gene set, a LASSO-regularized logistic regression model is used for training. The optimal regularization parameter is selected through 10-fold cross-validation, and genes with non-zero regression coefficients are identified as the core feature gene set.

5. The method according to claim 4, characterized in that, The loss function of the LASSO-regularized logistic regression model is: in: It is the first The true label of each sample; It is through the current model to the first Predicted values ​​for each sample; It is a regularization parameter that controls the intensity of feature selection; It is the first The regression coefficient of each gene; It is the total number of features; It represents the total number of samples.

6. The method according to claim 1, characterized in that, The core feature gene set includes multiple of the following genes: NF1, SF3B1, CUX1, CDKN2A, HOXA9, SPI1, MYD88, FLT3, STAT5B, ERG, PIK3CA, IDH1, WT1, MN1, RARA, CXCR4, DDX41, BCOR, HOXA10, DUX4, MCL1, CDKN2B, JAK1, CSF3R.

7. The method according to claim 1, characterized in that, Step S6 is followed by: Model saving steps: Save the model parameters of the trained risk probability prediction model, the list of core feature genes, and the performance evaluation results in a unified manner to ensure the traceability and reproducibility of the model.

8. The method according to claim 1, characterized in that, The training data in step S1 comes from multiple independent data sources, including at least the TARGET, TCGA-LAML, GTEx, and GSE227832 queues; the acute leukemia samples include acute myeloid leukemia samples and acute lymphoblastic leukemia samples.

9. A system for constructing an acute leukemia risk prediction model based on transcriptome data, characterized in that, include: The data acquisition module is used to acquire raw counting matrices of peripheral blood transcriptome sequencing from multiple sources and their corresponding sample labels; The data preprocessing module is used to perform cleaning, normalization, logarithmic transformation and batch effect correction on the original counting matrix to generate a standardized training representation feature matrix. The feature selection module is used to select the core feature gene set based on the corrected training expression feature matrix, using a multi-stage strategy that combines random forest and LASSO regularized logistic regression. The model training module is used to train an acute leukemia risk probability prediction model by taking the expression level of the core feature gene set as input, the sample label as the prediction target, and using a LASSO-regularized logistic regression algorithm and a cross-validation strategy. The model evaluation module is used to evaluate the performance of the trained risk probability prediction model using a test set and output evaluation metrics.

10. The system according to claim 9, characterized in that, The system is encapsulated in a server, cloud computing platform, or medical artificial intelligence inference terminal, and outputs the risk probability prediction model and its evaluation results through a graphical interface or application programming interface.

Images