Machine learning-based multiple myeloma risk prognosis layering model

Abstract

The invention discloses a multiple myeloma risk prognosis layering model based on machine learning. The method comprises the following steps: acquiring single cell sequencing data of a multiple myeloma patient and a healthy person, and screening out differential expression genes between the patient and the healthy person; obtaining transcriptome sequencing data of a patient, and screening out candidate genes, which are remarkably related to the total lifetime of the patient, in differentially expressed genes in the transcriptome sequencing data; the method comprises the steps of integrating transcriptome sequencing data into a transcriptome sequencing data set on the basis of candidate genes, introducing external transcriptome sequencing data as an external verification set, obtaining a prognosis layering model through algorithm combination matching, and testing the prediction performance of the prognosis layering model through tests including a column graph and an AUC curve. Based on the synergistic effect among multiple omics technologies such as single cell transcriptome sequencing, machine learning, transcriptome sequencing and the like, the model can quickly and accurately predict the prognosis condition of a patient in the early stage of a disease, so as to assist in clinically formulating a personalized treatment scheme.

Description

Technical Field

[0001] The present invention relates to a prognostic model for multiple myeloma, and in particular to a risk prognostic stratification model for multiple myeloma based on machine learning, belonging to the technical field of computer-assisted clinical analysis. Background Art

[0002] Multiple myeloma (MM) is a hematologic malignancy in which complex interactions between tumor and immune cells are essential for disease progression. However, the overall framework of immune cells within the bone marrow tumor microenvironment (BM-TME) and MM pathogenesis remains understudied. Monocytes, as a crucial component of the BM-TME, bridge the innate and adaptive immune responses and regulate the BM-TME through multiple mechanisms, including inducing immune tolerance, promoting angiogenesis, and facilitating tumor cell dissemination. Altered transcriptional signatures of monocytes during MM disease development and progression are associated with patient prognosis and response to immunotherapy. However, no studies have yet reported a link between monocyte genetic alterations and prognostic prediction in MM.

[0003] In the era of precision medicine, the Revised International Staging System (R-ISS), due to its limited parameters, cannot meet clinicians' needs for ideal prognostic indicators. Existing MM prognostic models often use single or biased algorithms or lack validation across multiple datasets, resulting in suboptimal performance or overfitting. Summary of the Invention

[0004] To address the challenges of existing technologies, the present invention aims to provide a machine learning-based risk and prognosis stratification model for multiple myeloma. This model, based on the synergy of single-cell transcriptome sequencing combined with machine learning and other multi-omics technologies, screens prognostic genes in multiple myeloma patients and constructs a prognostic prediction model. This model can quickly and accurately predict the prognosis of multiple myeloma patients in the early stages of the disease, thereby assisting clinicians in developing personalized treatment plans.

[0005] To achieve the above technical objectives, the present invention provides a risk prognosis stratification model for multiple myeloma based on machine learning, comprising:

[0006] Step S1, obtaining single-cell sequencing data of multiple myeloma patients and healthy subjects, and screening for differentially expressed genes between multiple myeloma patients and healthy subjects;

[0007] Step S2: obtaining transcriptome sequencing data of multiple myeloma patients, and screening candidate genes among the differentially expressed genes that are significantly associated with the patients' overall survival;

[0008] Step S3: Based on the candidate genes provided in step S2, the transcriptome sequencing data in step S2 are integrated into a transcriptome sequencing dataset, and external transcriptome sequencing data are introduced as an external validation set, and then a prognostic stratification model is obtained through algorithm combination matching;

[0009] Step S4: testing the predictive performance of the prognostic stratification model through tests including nomograms and AUC curves.

[0010] The multiple myeloma risk prognosis stratification model provided by the present invention can quickly assist clinicians in predicting the patient's prognosis in the early stages of the disease, providing a strong basis for timely adjustment of medication and formulation of individualized treatment plans.

[0011] As a preferred solution, the screening process for differentially expressed genes is as follows: after quality control cleaning and cell annotation of single-cell sequencing data, immune cells are identified, and then the differentially expressed genes derived from monocytes between MM patients and healthy people are identified through the pseudobulks package.

[0012] As a preferred solution, the screening process of the candidate genes is as follows: based on the transcriptome sequencing data of multiple myeloma patients, a single factor Cox regression analysis is performed on each differentially expressed gene.

[0013] As a preferred solution, the criterion for determining whether the differentially expressed genes are significantly correlated is: pvalue<0.05.

[0014] The gene screening process is one of the important steps of the present invention. This process combines single-cell sequencing and transcriptome sequencing data to more accurately focus on monocyte genes closely related to multiple myeloma, avoiding the gene screening bias that may exist in traditional single transcriptome sequencing data-based screening methods due to the complexity of cell types in the bone marrow.

[0015] As a preferred solution, the transcriptome sequencing dataset is divided into a training set and a test set at a data volume ratio of 6 to 8:2 to 4. Further preferably, the transcriptome sequencing dataset is divided into a training set and a test set at a data volume ratio of 7 to 3.

[0016] As a preferred solution, the algorithm combination matching process is: each classic machine learning algorithm is matched in the training set and the validation set through a leave-one-out cross-validation framework.

[0017] This method utilizes a leave-one-out cross-validation framework for algorithm combination matching, fully utilizing transcriptome sequencing datasets. Only one transcriptome sequencing dataset is retained for validation at a time, while the remaining datasets are used for training. This effectively addresses the issue of small sample sizes in medical data. The model provides a nearly unbiased estimate of model generalization error, accurately assessing model performance and ensuring the reliability and effectiveness of the selected algorithm combination in practical applications. Furthermore, by continuously adjusting and validating model parameters, this method effectively avoids overfitting, ensuring the model's stable predictive performance on new data, and enhancing the practical application value of the prognostic stratification model for multiple myeloma.

[0018] As a preferred solution, the classic machine learning algorithms include LASSO algorithm, Stepwise Cox algorithm, RSF algorithm, CoxBoost algorithm, Enet algorithm, GBM algorithm, Survival-SVM algorithm, SuperPC algorithm, Ridge algorithm and plsRcox algorithm.

[0019] As a preferred solution, the process of obtaining the prognostic stratification model is as follows: by calculating the consistency index C-index of all matching algorithm combinations, selecting the algorithm combination corresponding to the highest value, and obtaining the prognostic stratification model.

[0020] As a preferred solution, the formula for calculating the risk score of the monocyte gene prognostic signature (MGPS) in the prognostic stratification model is:

[0021] Formula 1: MGPS risk score = (C1×PFDN2) + (C2×TUBA1B) + (C3×ERH) + (C4×FAM49B) + (C5×HNRNPR) + (C6×RHOC) + (C7×ANAPC11) + (C8×IL32) + (C9×GLIPR2) + (C10×ZNF90) + (C11×YWHAZ) - (C12×PECAM1) - (C13×ENY2) - (C14×PSAP) - (C15×SNRPC) - (C16×IL16) - (C1×PCGF5) - (C18×ERP29);

[0022] In formula 1, C1~C18 are the scoring coefficients of each gene and are constants.

[0023] As a preferred solution, the predictive performance detection process of the prognostic stratification model includes: dividing the patients in the transcriptome sequencing data set into high-risk group and low-risk group according to the MGPS risk score, performing Kaplan-Meier survival analysis on each group, and calculating the AUC of the model prediction results.

[0024] As a preferred solution, the classification standard for the high-risk group and the low-risk group is one of the median value, weighted average value, Euclidean distance average value, and Mahalanobis distance average value. Further preferably, the classification standard for the high-risk group and the low-risk group is the median value.

[0025] Compared with the prior art, the beneficial technical effects of the technical solution provided by the present invention are:

[0026] 1) The model provided by this invention is based on the synergy between single-cell transcriptome sequencing, machine learning, and multi-omics technologies such as transcriptome sequencing. It screens prognostic genes for multiple myeloma patients and constructs a prognostic prediction model. It can quickly and accurately predict the prognosis of multiple myeloma patients in the early stages of the disease, thereby assisting clinicians in formulating personalized treatment plans.

[0027] 2) The technical solution provided by the present invention utilizes machine learning and multi-omics sequencing technologies, and through leave-one-out cross-validation, it traverses the combination matching of classic machine learning algorithms. Combined with a large amount of patient data from multiple sources, it fundamentally solves the problems of insufficient accuracy and poor reliability caused by the overly single algorithm selection, bias, and lack of data verification in the prognostic models in the existing technology. After testing, this prognostic stratification model is an independent prognostic factor for patients with multiple myeloma (HR 2.72, 95% CI: 1.84-4.0, p<0.001), and its AUC for 1-4 years shows reliable predictive performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of the modeling process of the prognostic stratification model provided in Example 1 of the present invention;

[0029] in, Figure 1 (A) UMAP diagram showing the distribution of 12 cell types annotated by single-cell sequencing data. Figure 1 (B) A volcano plot showing 1447 differentially expressed genes in monocytes between MM patients and healthy controls. Figure 1 (C) is the model consistency index C-index diagram of 101 machine learning algorithm matching combinations in the training set and validation set. Figure 1 (D) is a bar graph showing the component genes of the MGPS model and the weighted coefficient coef of each gene;

[0030] Figure 2 This is a graph showing the predictive performance of the prognostic stratification model provided in Example 1 of the present invention;

[0031] in, Figure 2 (A) is the survival curve of MGPS high and low risk groups (weighted average) of 590 MM patients in the MMRF training set. Figure 2 (B) is the survival curve of MGPS high and low risk groups (based on median) of 590 MM patients in the MMRF training set. Figure 2 (C) is the survival curve of MGPS high- and low-risk groups of 254 MM patients in the MMRF validation set. Figure 2 (D) is the survival curve of MGPS high- and low-risk groups of 426 MM patients in the GSE136337 validation set. Figure 2 (E) Comparison of the number of deaths and survivors in the MGPS high-risk and low-risk groups of 549 MM patients in the GSE24080 validation set at the 720th and 1440th days of follow-up. Figure 2 (F) is the receiver operating characteristic (ROC) curve of 590 MM patients in the MMRF training set at 1, 2, 3, and 4 years. Figure 2 (G) is the receiver operating characteristic (ROC) curve of 254 MM patients in the MMRF validation set at 1, 2, 3, and 4 years. Figure 2 (H) is the receiver operating characteristic (ROC) curve of 426 MM patients in the GSE136337 validation set at 1, 2, 3, and 4 years. Figure 2 (I) Forest plot of univariate Cox regression analysis of MGPS and clinical covariates in 426 MM patients in the GSE136337 validation set. DETAILED DESCRIPTION

[0032] To facilitate understanding of the present invention, the present invention will be described in more detail below with reference to the accompanying drawings. It should be noted that the embodiments described are only some embodiments of the present invention and not all embodiments. All other embodiments derived by those skilled in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the invention.

[0033] Example 1

[0034] This embodiment provides a risk prognosis stratification model for multiple myeloma based on machine learning, the specific process of which is as follows:

[0035] 1) Single-cell sequencing data GSE124310 of MM patients (n=23) and healthy subjects (n=9) were obtained from the GEO database. After quality control cleaning and cell annotation, 12 types of immune cells were identified, such as Figure 1 (A) As shown; 1447 differentially expressed genes (moDEGs) derived from monocytes between MM patients and healthy subjects were identified using the pseudobulks package, as shown Figure 1 (B) shown;

[0036] 2) Using transcriptome sequencing data from 844 newly diagnosed multiple myeloma patients (MMRF) with survival information, univariate Cox regression analysis was performed on 1,447 genes. Based on a p-value < 0.05, 482 candidate genes were identified that were significantly associated with overall survival (OS) in MM patients. Of these, 324 genes had a hazard ratio (HR) > 1, and 158 genes had a hazard ratio (HR) < 1.

[0037] 3) The 844 patients in the MMRF cohort were divided into a training set (n=590) and a validation set (n=254) in a 7:3 ratio. There were no significant differences in age, sex, or survival between the training set and the validation set. Two transcriptome sequencing data sets, GSE136337 (426 MM patients) and GSE24080 (549 MM patients), were included as external validation sets.

[0038] 4) Ten machine learning algorithms: LASSO algorithm, Stepwise Cox algorithm, RSF algorithm, CoxBoost algorithm, Enet algorithm, GBM algorithm, Survival-SVM algorithm, SuperPC algorithm, Ridge algorithm and plsRcox algorithm were matched with 101 algorithm combinations based on the leave-one-out cross-validation framework in the training set. The prediction performance of all models was evaluated by calculating the model C-index of the above 101 algorithm combinations in all training sets and validation sets. The results are as follows: Figure 1 As shown in (C), the CoxBoost+RSF model has the highest C-index of 0.7, so the CoxBoost+RSF model is the prognostic stratification model;

[0039] The risk score calculation formula of the monocyte-related gene prognostic signature (MGPS) in the prognostic stratification model is:

[0040] Formula 1: MGPS risk score = (0.0582 ×PFDN2) + (0.2206 ×TUBA1B) +(0.1787 ×ERH) + (0.0483 ×FAM49B) +(0.1253 ×HNRNPR) +(0.1223 ×RHOC) +(0.2697 ×ANAPC11) +(0.0369 ×IL32) +(0.1522 ×GLIPR2) +(0.0723 ×ZNF90) +(0.1402 ×YWHAZ) -(0.1041 ×PECAM1) - (0.0101 ×ENY2) - (0.2086 ×PSAP) - (0.0086 ×SNRPC) -(0.2099 ×IL16) - (0.1319 ×PCGF5) - (0.2628 ×ERP29);

[0041] 5) To test the predictive performance of the above-mentioned prognostic stratification model, MM patients in multiple training sets and validation sets were divided into high-risk and low-risk groups according to the median of the MGPS risk score, and Kaplan-Meier survival analysis was performed. The results are shown in Figure 5. Figure 2 (A)~ Figure 2 As shown in (E), the figure shows that the OS of the high-risk group in all cohorts was significantly worse than that of the low-risk group. Further, the MGPS risk score was tested for 1, 2, 3, and 4 years in the training set, validation set, and two external validation sets. The results are shown in Figure 5. Figure 2 (F)~ Figure 2 As shown in (H), the ROC curve of the MGPS risk score shows a robust and reliable predictive performance in predicting the 1- to 4-year AUC of MM patients, making it an ideal prognostic indicator.

[0042] Furthermore, in order to further illustrate the excellent prediction accuracy of the model provided in this embodiment, the present invention also performed a single factor cox regression analysis forest plot of the MGPS risk score and clinical covariates on the above model in the GSE136337 validation set, and the results are shown as follows: Figure 2 As shown in (I), the figure shows that the MGPS risk score and clinical covariates such as age and β2m in this model are independent prognostic factors for MM patients, and the MGPS risk score (HR 2.72, 95%CI: 1.84-4.0, p<0.001) has a higher hazard ratio. Compared with other clinical covariates, it is more sensitive to the prediction of MM patients and has higher prediction accuracy.

[0043] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should also be regarded as within the scope of protection of the present invention.

Claims

1. A risk prognosis stratification model for multiple myeloma based on machine learning, characterized in that: include: Step S1, obtaining single-cell sequencing data of multiple myeloma patients and healthy subjects, and screening for differentially expressed genes between multiple myeloma patients and healthy subjects; Step S2: obtaining transcriptome sequencing data of multiple myeloma patients, and screening candidate genes among the differentially expressed genes that are significantly associated with the patients' overall survival; Step S3: Based on the candidate genes provided in step S2, the transcriptome sequencing data in step S2 are integrated into a transcriptome sequencing dataset, and external transcriptome sequencing data are introduced as an external validation set, and then a prognostic stratification model is obtained through algorithm combination matching; Step S4: testing the predictive performance of the prognostic stratification model through tests including nomograms and AUC curves.

2. The multiple myeloma risk and prognosis stratification model based on machine learning according to claim 1, characterized in that: The screening process of the differentially expressed genes is as follows: after quality control cleaning and cell annotation of the single-cell sequencing data, immune cells are identified, and then the differentially expressed genes of monocyte origin between MM patients and healthy people are identified through the pseudobulks package.

3. The multiple myeloma risk and prognosis stratification model based on machine learning according to claim 1, characterized in that: The screening process of the candidate genes is as follows: based on the transcriptome sequencing data of multiple myeloma patients, a single factor Cox regression analysis is performed on the differentially expressed genes; the criterion for determining whether the differentially expressed genes are significantly correlated is: pvalue<0.

05.

4. The multiple myeloma risk and prognosis stratification model based on machine learning according to claim 1, characterized in that: The transcriptome sequencing dataset was divided into a training set and a test set according to a data volume ratio of 6-8:2-4.

5. The multiple myeloma risk and prognosis stratification model based on machine learning according to claim 1, characterized in that: The algorithm combination matching process is: each classic machine learning algorithm is matched in the training set and the validation set through a leave-one-out cross-validation framework.

6. The multiple myeloma risk and prognosis stratification model based on machine learning according to claim 1 or 5, characterized in that: The classic machine learning algorithms include LASSO algorithm, Stepwise Cox algorithm, RSF algorithm, CoxBoost algorithm, Enet algorithm, GBM algorithm, Survival-SVM algorithm, SuperPC algorithm, Ridge algorithm and plsRcox algorithm.

7. The multiple myeloma risk and prognosis stratification model based on machine learning according to claim 1, characterized in that: The process of obtaining the prognostic stratification model is as follows: by calculating the consistency index C-index of all matching algorithm combinations, selecting the algorithm combination corresponding to the highest value, and obtaining the prognostic stratification model.

8. The multiple myeloma risk and prognosis stratification model based on machine learning according to claim 1, characterized in that: The formula for calculating the risk score of the monocyte gene prognostic signature (MGPS) in the prognostic stratification model is: Formula 1: MGPS Risk Score = (C1 × PFDN2) + (C2 × TUBA1B) + (C3 × ERH) + (C4 × FAM49B) + (C5 × HNRNPR) + (C6 × RHOC) + (C7 × ANAPC11) + (C8 × IL32) + (C9 × GLIPR2) + (C10 × ZNF90) + (C11 × YWHAZ) - (C12 × PECAM1) - (C13 × ENY2) - (C14 × PSAP) - (C15 × SNRPC) - (C16 × IL16) - (C1 × PCGF5) - (C18 × ERP29) ; In formula 1, C1~C18 are the scoring coefficients of each gene and are constants.

9. The multiple myeloma risk and prognosis stratification model based on machine learning according to claim 1, characterized in that: The predictive performance testing process of the prognostic stratification model includes: dividing patients in the transcriptome sequencing data set into a high-risk group and a low-risk group according to the MGPS risk score, performing Kaplan-Meier survival analysis on each group, and calculating the AUC of the model prediction results.

10. The multiple myeloma risk and prognosis stratification model based on machine learning according to claim 9, characterized in that: The classification standard of the high-risk group and the low-risk group is one of the median value, weighted average value, Euclidean distance average value and Mahalanobis distance average value.

Images