A method for constructing a multi-cancer risk prediction model based on PBMC single-cell sequencing and application thereof

Abstract

The application discloses a kind of based on PBMC single-cell sequencing multi-cancer risk prediction model construction method and its application, belong to biotechnology field.The method is by integrating multi-source single-cell data, obtains high-quality immune cell atlas by quality control, batch correction and cell annotation, finally by self-test single-cell data queue is verified.Pseudo batch analysis and differential expression screening are used to obtain cell characteristic genes consistent across cancer species, and then LASSO regression dimensionality reduction is performed to finally determine a set of 39 key characteristic genes.Based on this, the screening model exhibits high accuracy (92.2%-97.3%) in both training and independent validation, and can evaluate the risk of 15 cancers at once.The model only needs a small amount of peripheral blood, adapts to conventional commercial scRNA-seq platform, has the advantages of non-invasive, high generalization, strong specificity and good interpretability, and is suitable for large-scale early screening in clinical practice, and has clear conversion prospect.

Description

Technical Field

[0001] This invention belongs to the field of biotechnology and relates to a method for constructing a multi-cancer risk prediction model based on PBMC single-cell sequencing and its application. Background Technology

[0002] Cancer is a major global public health problem, and early detection is one of the important means to reduce mortality, improve survival rates, and improve prognosis. Theoretically, simultaneous early screening for multiple common cancers can maximize the advancement of the treatment window, achieving "one screening, multiple benefits," making it a highly promising public health strategy. Multi-cancer combined screening is an effective strategy to reduce cancer mortality and extend the treatment window. The current mainstream cancer screening model is a "single cancer, single method" approach, such as using low-dose spiral CT to screen for lung cancer, mammography to screen for breast cancer, and colonoscopy to screen for colorectal cancer. This model has the following inherent drawbacks: Severely insufficient coverage. Only a few types of cancer have widely accepted, population-based screening programs, while for many common or high-mortality malignancies such as stomach cancer, liver cancer, pancreatic cancer, and esophageal cancer, there is a lack of effective early screening methods; Poor accessibility and adherence. Existing screening methods are often invasive (e.g., colonoscopy), involve radiation exposure (e.g., CT), or are inconvenient and costly, resulting in low public participation and difficulty in large-scale implementation; Low overall efficiency. Conducting individual cancer screenings one by one consumes a large amount of medical resources and individual time, and is not cost-effective in a public health sense.

[0003] To overcome the aforementioned bottlenecks, early detection technologies for multiple cancer types based on liquid biopsy have become a research frontier. Among these, detection using the methylation or mutation characteristics of circulating tumor DNA (ctDNA) in blood has made some progress. However, ctDNA has extremely low abundance in the blood of early-stage cancer patients, requiring extremely high detection sensitivity, and the accuracy of tissue tracing still needs improvement. These factors limit its clinical translation and application effectiveness. Peripheral blood, as a carrier of systemic immune responses, contains rich information reflecting the body's pathological state. The occurrence and development of tumors trigger systemic immune system remodeling, and these changes are reflected in the composition, proportion, activation, and functional state of peripheral blood immune cells. Therefore, utilizing peripheral blood transcriptome information for cancer detection is considered a highly promising complementary approach.

[0004] Existing Bulk RNA-seq (population RNA sequencing) data analysis models based on peripheral blood mononuclear cells (PBMCs) suffer from fundamental technical flaws, making it difficult to meet the requirements of high-precision screening in terms of predictive efficacy. Bulk RNA-seq technology sequences millions of heterogeneous PBMCs (including various functional subpopulations such as T cells, B cells, NK cells, and monocytes), resulting in gene expression data that is the average of transcripts from all cells. This method leads to severe loss and dilution of key biological information: 1) Loss of cellular heterogeneity information: It cannot distinguish specific immune cell subpopulations that may be most relevant to the early cancer response (such as a specific subtype of effector T cells or dendritic cells); 2) Submergence of weak specific signals: Weak but specific gene expression changes triggered by early cancer, which may only exist in a few cell subpopulations, are masked by background noise from a large number of other unrelated cells; 3) Lack of biological mechanism analysis: Models built based on mixed data are like "black boxes," unable to reveal the specific cellular origin and functional state changes of the cancer immune response, resulting in poor model interpretability and a lack of clear direction for further optimization.

[0005] In summary, there is an urgent need for a novel screening method and model that can overcome the limitations of Bulk RNA-seq technology, fully utilize the high-dimensional and detailed information of peripheral blood immune profiles, thereby achieving high sensitivity and specificity, and enabling early risk assessment of various cancers. Summary of the Invention

[0006] To address the technical problem that existing multi-cancer screening technologies based on peripheral blood Bulk RNA-seq cannot resolve immune cell heterogeneity due to the mixed sequencing principle, resulting in the dilution and loss of specific immune signals related to early cancer, leading to low model prediction efficacy, poor interpretability, and difficulty in achieving high-precision and high-coverage early cancer screening, the present invention aims to provide a method for constructing and applying a multi-cancer risk prediction model based on PBMC single-cell sequencing.

[0007] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, the present invention provides a method for constructing a multi-cancer risk prediction model based on PBMC single-cell sequencing, comprising: S1: Obtain single-cell sequencing data of PBMCs from various cancer patients and healthy controls to construct a pan-cancer PBMC single-cell atlas; S2, Based on the pan-cancer PBMC single-cell atlas, the data of each immune cell subset are integrated to construct a pseudo-batch expression matrix at the sample-cell subset level; S3, for each immune cell subset, the pseudo-batch expression matrix is ​​used to perform differential expression analysis between patient samples and healthy control samples of each cancer type to obtain the preliminary differentially expressed genes of each immune cell subset in each cancer type. S4, perform cross-cancer screening on the preliminarily differentially expressed genes to obtain a candidate gene set; S5. The candidate gene set is screened using the LASSO regression algorithm to obtain the expression data of the core characteristic gene set of each immune cell subset and its corresponding regression coefficient. S6, based on the expression data of the core feature gene set and its corresponding regression coefficients, calculates the risk score of each sample in each immune cell subpopulation to form a multidimensional risk feature vector. The multidimensional risk feature vector of the sample is used as input and trained using machine learning algorithms to obtain a multi-cancer risk prediction model.

[0008] In step S1, the various cancers include at least 15 types. The PBMC single-cell transcriptome sequencing data are derived from at least 36 datasets and, after quality control, contain at least 305 tumor patient samples.

[0009] In step S2, the immune cell subsets include B cells and CD14 cells. + Monocytes, CD16 + Monocytes, CD4 + Conventional T cells, CD8 + There are 11 cell types, including T cells, dendritic cells, natural killer cells, plasma cells, platelets, proliferating T cells, and regulatory T cells.

[0010] In step S4, the cross-cancer screening includes: screening genes with consistent expression trends in at least a preset proportion of cancer types, wherein consistent expression trends refer to both being upregulated or both being downregulated; and performing secondary screening based on the expression percentage of the screened genes in cancer types and healthy controls to obtain a candidate gene set.

[0011] Preferably, the preset ratio is 60%, and the secondary screening includes: for upregulated genes, the expression percentage must be higher than 0.3 in at least 60% of cancer types and lower than 0.3 in the healthy control group; for downregulated genes, the expression percentage must be lower than 0.3 in at least 60% of cancer types and higher than 0.3 in the healthy control group.

[0012] In step S5, the core feature gene set contains 39 genes.

[0013] Preferably, the risk score is calculated using the following formula: RiskScore i,c =∑ g∈Gc (β g,c × X i,g,c ) in, RiskScore i,c Indicates sample i and cell type c Risk score, Gc Cell type c The set of feature genes selected by the LASSO model β g,c For genes g In cell type c LASSO regression coefficients in X i,g,c For the sample i In cell type c Pseudo-batch expression value of gene g in the middle.

[0014] The machine learning algorithm in step S6 includes at least one of random forest, support vector machine, XGBoost, neural network, or a combination thereof.

[0015] Secondly, this invention provides a multi-cancer risk prediction system, constructed using the method described above, comprising: The data acquisition and atlas construction module is used to acquire single-cell sequencing data of PBMCs from various cancer patients and construct single-cell atlases of PBMCs from tumor patients. The pseudo-batch matrix construction module is used to integrate the data of each immune cell subset and construct a pseudo-batch expression matrix at the sample-cell subset level. The feature gene screening module is used to perform differential expression analysis of each immune cell subset across cancer types and obtain a candidate gene set; The core feature screening module is used to screen the candidate gene set using the LASSO regression algorithm to obtain the core feature gene set and its corresponding regression coefficients. The risk score calculation module is used to calculate the cell type-specific risk score for each sample based on the core feature gene set and its regression coefficients, forming a multidimensional risk feature vector. The model training and evaluation module is used to train and validate the multi-dimensional risk feature vectors through machine learning algorithms to obtain a multi-cancer risk prediction model. The risk assessment module is used to receive PBMC single-cell sequencing data from the subject and output the risk assessment results of the subject having multiple cancers based on the multi-cancer risk prediction model.

[0016] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for constructing a multi-cancer risk prediction model based on PBMC single-cell sequencing.

[0017] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for constructing a multi-cancer risk prediction model based on PBMC single-cell sequencing.

[0018] Fifthly, the present invention provides the use of the multi-cancer risk prediction model constructed by the method or the system described herein in the preparation of reagent kits or diagnostic devices for early cancer screening or risk assessment.

[0019] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a method for constructing a multi-cancer risk prediction model based on PBMC single-cell sequencing. This method solves the core technical problems of signal dilution, loss of heterogeneity, and poor generalization ability of existing Bulk RNA-seq technology by using a complete technical chain: single-cell resolution analysis of immune heterogeneity → pseudo-batch integration to eliminate single-cell data sparsity → cross-cancer screening to extract common features → LASSO dimensionality reduction to obtain core genes → cell type-specific risk score modeling. This results in a multi-cancer risk prediction model with high accuracy (AUC>0.94), strong generalization ability (covering 15 types of cancer), and good interpretability (traceable to specific immune cell subpopulations). The data used to build the model comes from a variety of representative cancers, so the model can monitor multiple cancers simultaneously with a single blood sample. Through steps S4 (cross-cancer consistency screening) and S5 (LASSO screening), the core features that truly reflect the common immune response of tumors are separated from the differentially expressed genes specific to each cancer type. This effectively solves the problems of high noise in single-cell data, strong interference from cancer type specificity, and the curse of feature dimensionality. The univariate (gene expression) is transformed into a multivariate (cell type risk score), providing higher quality and higher-level feature input for the final ensemble learner.

[0020] Furthermore, the study includes at least 15 different cancers, covering both common cancers (lung cancer, breast cancer, colorectal cancer, etc.) and cancers difficult to screen early (pancreatic cancer, liver cancer, etc.), demonstrating broad clinical applicability. Through cell type modeling, it captures the functional differences of different immune cells in tumor immune responses (such as the immunosuppression of Treg cells, CD8+ cells...). (T cell cytotoxic activity); decouples signals that were originally mixed in Bulk analysis into independent feature dimensions, avoiding the drowning out of key signals. The prediction results can be traced back to specific immune cell subpopulations, transforming the "black box model" into an "interpretable model." Cross-cancer screening enables the model to accurately predict new cancer types not present in the training set. The screened genes are enriched in core immune response pathways (T cell activation, interferon response, etc.) and have clear biological significance. The number of genes initially differentially expressed has been reduced from hundreds to 39, a dimensionality reduction of >80%, solving the overfitting risk of high-dimensional feature modeling. The number of 39 genes is fully compatible with qPCR or multi-gene chip platforms, reducing the detection cost from thousands of yuan for single-cell sequencing to hundreds of yuan, achieving faster and more economical results. The constructed model has the advantage of being non-invasive and can be combined with other physical examination items, making it easily accepted by subjects. This model supports multi-cancer risk assessment based on a single peripheral blood sample.

[0021] The system provided by this invention automates the entire process from data input to risk assessment. Each module is functionally independent, which facilitates system maintenance, upgrades, and individual optimization of certain functions. Users only need to input PBMC single-cell sequencing data, and the system can automatically output risk assessment results without the need for manual intervention in complex analysis processes.

[0022] The application provided by this invention transforms high-precision prediction models into practical clinical testing tools, achieving breakthroughs in five dimensions: testing performance, coverage, ease of operation, cost control, and result interpretability. It possesses the mature conditions to move from laboratory to large-scale clinical application and can be prepared into corresponding reagent kits or diagnostic devices, which can be widely used in medical scenarios such as health management centers, laboratories, and oncology departments. Attached Figure Description

[0023] Figure 1 It is the process of collecting single-cell sequencing data of pan-cancer PBMCs to construct a pan-cancer atlas. Among them, A is the distribution map of cell number in different cancers, B is cell type annotation and UMAP visualization, and C is cell annotation of each immune cell population using marker genes. Figure 2 To predict the ROC score of the model on the test and validation sets of public single-cell data, where A is the ROC curve of the training set and B is the ROC curve of the test set.

[0024] Figure 3This represents the detection performance of the prediction model in the self-tested single-cell dataset, where A is the grouping information of the self-tested data queue, and B is the ROC curve of the self-tested single-cell transcriptome data. Detailed Implementation

[0025] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. These descriptions are intended to explain the invention and not to limit it.

[0026] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0027] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention 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 the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0028] In this article, the term "PBMC" stands for human peripheral blood mononuclear cells.

[0029] The use of patient samples involved in this invention has been approved by the Medical Ethics Committee of the First Affiliated Hospital of Xi'an Jiaotong University (ID: LLSBPJ-2024-448), and all patients have signed informed consent forms.

[0030] Example 1: Construction and Validation of a Multi-Cancer Risk Prediction Model Based on PBMC Single-Cell Sequencing This invention aims to construct a non-invasive pan-cancer early risk prediction model based on peripheral blood immune status using single-cell transcriptome data from peripheral blood mononuclear cells (PBMCs) of healthy individuals and patients with multiple cancer types. The specific steps are as follows: (1) Sample source and data acquisition A total of 36 single-cell PBMC transcriptome datasets were downloaded from public databases including the National Center for Biotechnology Information (GEO), the European Nucleotide Archive (ENA), the Genome Sequence Archive (GSA), and the National Omics Data Encyclopedia (NODE). A total of 329 samples were collected from 15 different cancer types (lung cancer, breast cancer, colorectal cancer, liver cancer, stomach cancer, esophageal cancer, pancreatic cancer, ovarian cancer, prostate cancer, bladder cancer, kidney cancer, melanoma, thyroid cancer, head and neck cancer, and lymphoma) as shown in Table 1. Figure 1 (A)

[0031] Table 1: Datasets of 15 different cancer types

[0032] (2) Data preprocessing and batch correction Data preprocessing and analysis were performed using R (version 4.4.2) and the Seurat package (version 5.2.1). First, genes detected in at least 90% of the datasets (≥32 genes) were selected, resulting in a common gene set of 17,357 genes. Quality control criteria were: 500-8000 detected genes, mitochondrial gene proportion <20%, hemoglobin gene proportion <5%, and samples with >800 cells retained. After quality control, 305 samples, totaling 1,858,536 high-quality cells, were finally included.

[0033] To assess and label the sources of batch effects, we first independently determined the existence of significant batch effects within each dataset. Standard preprocessing procedures were performed on each dataset, including normalization, screening for highly variable genes, scaling, principal component analysis (PCA), and UMAP dimensionality reduction visualization. The group distributions were then displayed in both PCA and UMAP spaces in conjunction with sample labels. If samples showed significant separation in the low-dimensional space, the dataset was considered to have internal batch effects, and the sample names were assigned as batch labels. If samples highly overlapped, no significant internal batch differences were considered, and the entire dataset was treated as a single batch. After setting batch labels for all datasets, they were merged and used as input for subsequent batch effect removal. In a Python 3.10 environment, the omicverse package (version 1.6.10) was used to perform batch effect correction on all data according to batch labels using the Harmony method.

[0034] (3) Cell type annotation, pseudo-batch integration and differential expression analysis Cell type annotation was performed by retrieving marker genes for each cell type through literature search and then annotating them based on their expression distribution across different cell populations. The cell type corresponding to the marker gene with the highest expression level was labeled as the cell type of that cell population. The marker genes used are as follows: circulating tumor cells (CTCs) – CD63, KRT18, KRT8; T cells – CD3D, CD3G; CD4… + T cells (CD4) + T) is CD4, IL7R, TRBC2; CD8 + T cells (CD8) + T cells are CD8A, CD8B, GZMK, GZMA, CCL5, GZMB, GZMH; regulatory T cells (Treg cells) are FOXP3, IL2RA, IKZF2; proliferating cells are MKI67, TOP2A; natural killer (NK) cells are GNLY, NKG7, CD247, FCERT1G, TYROBP, KLRG1, KLRF1; B cells are MS4A1, PAX5, BLK, IGHD, IGHM; plasma cells are MZB1, HSP90B1, FNDC3B, PRDM1, IGKC, JCHAIN; dendritic cells (DCs) are LYZ, CD1C, GZMB, IL3RA, COBLL1, TCF4; CD14 + Monocytes (CD14) + Mono) is FCN1, CD14; CD16 + Monocytes (CD16+ Mono) were identified as TCF7L2, FCGR3A, and LYN; platelets were identified as PPBP and PF4. Cell populations were annotated based on known marker genes, identifying B cells and CD14+ cells. + Monocytes, CD16 + Monocytes, CD4 + Conventional T cells, CD8 + Subpopulations include T cells, dendritic cells, natural killer cells, plasma cells, platelets, proliferating T cells, regulatory T cells, and circulating tumor cells (CTC).

[0035] Differential expression analysis was conducted based on single-cell data. The screening criteria were: a) upregulated genes (log2FC > 1) or downregulated genes (log2FC < -1), and b) corrected Q value < 0.05. Differential expression analysis was performed on each immune cell subset between tumor patients and healthy controls to obtain preliminary differentially expressed genes for each subset in each cancer type (see attached). Figure 1 (As shown in B and C).

[0036] (4) Screening for common characteristic genes of multiple cancer types To screen for cell type-specific differentially expressed genes that are stable across multiple cancer types, and because circulating tumor cells (CTCs) have patient-specific mutations and are unsuitable for feature transfer between different cancer patients, the analysis was limited to all immune cell populations other than CTCs, specifically including B cells and CD14. + Monocytes, CD16 + Monocytes, CD4 + Conventional T cells (CD4 Tconv cells), CD8 + T cells, dendritic cells, natural killer cells, plasma cells, platelets, proliferating T cells, and regulatory T cells. The following three-step analysis workflow is used: a) Cross-cancer concordance screening By reviewing the differential expression analysis results of all cell subpopulations, genes with consistent expression trends (both upregulated or both downregulated) in at least 60% of cancer types (i.e., at least 9 of the 15 cancer types covered by this invention) were screened. b) Calculation of gene expression percentage For each candidate gene that passed the initial screening, the percentage of expression of each candidate gene in all 16 experimental groups (15 cancer types and 1 healthy control group) was calculated (i.e. the proportion of cells in the corresponding cell subpopulation that expressed the gene). c) Strict threshold secondary screening More stringent conditions were applied to the differentially expressed genes initially identified for secondary screening. Specifically, upregulated genes were required to have an expression percentage higher than 0.3% in at least 9 cancer types and an expression percentage lower than 0.3% in the healthy control group; downregulated genes were required to have an expression percentage lower than 0.3% in at least 9 cancer types and an expression percentage higher than 0.3% in the healthy control group.

[0037] Through the above screening process, a set of candidate characteristic genes that stably change in a multi-cancer background were finally obtained. Differential expression genes were preliminarily identified through differential analysis. The contribution of different immune cell subsets to these differentially expressed genes varied (e.g., ...). Figure 1The distribution and marker gene expression of different cell subpopulations are shown in Tables B and C. Regulatory T cells (Treg cells) and proliferating T cells contributed the most differentially expressed genes (102 and 67 respectively), suggesting they may play a central role in the pan-cancer immune response. Using the health status of the samples (tumor vs. healthy) as the response variable, LASSO regression analysis was performed on the candidate gene sets of the 11 immune cell subpopulations to further remove collinear genes and obtain the core characteristic genes and their regression coefficients for each cell subpopulation. Finally, 39 core characteristic genes were jointly screened from the 11 immune cell types (see Table 2), among which the upregulated gene was HLA- The following 11 genes were downregulated: DRB5, IGLC2, LIME1, PTPRCAP, RNASEK, ALDOA, ISG20, CXCR3, HPGD, SYTL3, and ATP6V0C. The following 28 genes were downregulated: NME2, FOSB, GNLY, ALYREF, PUS7, ID2, IGHG4, SNHG25, BHLHE40, EBP, PPP1R15A, PTMS, TRGC1, DUSP6, ITM2C, TNF, TRGC2, CCL4, GZMB, ITPR2, SOS1, HLA-DPA, IFNG, JUN, TRAPPC5, XIST, ATP3, and GZMH.

[0038] Table 2: 39 key genes identified through joint screening of 11 immune cell types

[0039] This screening process effectively eliminated noise and cancer-specific signals by combining "cross-cancer consistency screening" with "strict threshold screening for expression percentage". Ultimately, it obtained a set of core immune feature genes with stable and significant change patterns in different cancers, laying the foundation for the subsequent construction of a robust multi-cancer joint screening model.

[0040] (5) Risk score calculation For the sample i and cell type c Calculate the risk score using the following formula: RiskScore i,c =∑ g∈Gc ( β g,c × X i,g,c ) in, RiskScore i,c Indicates sample i and cell type c Risk score, Gc This is the set of core characteristic genes for cell type c. β g,c The LASSO regression coefficient of gene g in cell type c. X i,g,c For the sample i In cell type c Zhonggen g The pseudo-batch representation value (TPM) is obtained for each sample. An 11-dimensional risk feature vector is obtained for each sample.

[0041] (6) Integration model construction and validation All samples were divided into a 60% training set (189 cases) and a 40% validation set (115 cases). Nearly one hundred machine learning algorithms, including random forests, support vector machines, and gradient boosting machines, were systematically evaluated and compared on the training set, ultimately training an ensemble multi-cancer risk prediction model. The model was then evaluated using cross-validation and independent validation sets, with the optimal model selected through 5-fold cross-validation.

[0042] Model performance validation: The model demonstrates excellent discriminative ability on both the training and independent test sets. Specific performance metrics are as follows: Figure 2 The ROC curve is shown.

[0043] From the appendix Figure 2 The data shows that the model exhibits near-perfect discrimination on the test set (containing 189 samples) and validation set (containing 115 samples), with areas under the curve (AUC) of 0.959 (95% confidence interval: 0.945-0.973) and 0.942 (95% confidence interval: 0.922-0.962), respectively. This indicates that the model can very accurately identify cancer patient samples from healthy controls. This result confirms the model's effectiveness, rather than overfitting to the training data. The multi-cancer joint screening model constructed in this invention, based on specific immune characteristic genes screened from PBMC single-cell data, achieved excellent performance with an AUC > 0.94 in independent validation. This demonstrates that the model has a strong cancer risk identification capability, providing a reliable technical foundation for non-invasive, multi-cancer early screening based on peripheral blood immune profiles.

[0044] Example 2: Implementation of a Multi-Cancer Risk Prediction System Based on the model constructed in Example 1, a multi-cancer risk prediction system was developed, including the following modules: The data acquisition and atlas construction module uses Python 3.10 to write the data interface, supporting automatic download of PBMC single-cell sequencing data from GEO, GSA, ENA, and NODE databases, or receiving local data uploaded by users (supporting mainstream platform formats such as 10xGenomics and BD Rhapsody). It includes built-in quality control processes (500-8000 genes, mitochondrial gene ratio <20%, hemoglobin gene ratio <5%) and the Harmony batch correction algorithm, automatically constructing pan-cancer PBMC single-cell atlases.

[0045] The pseudo-batch matrix construction module integrates single-cell expression data of the same sample and cell type based on cell type annotation results using PseudoBulk, and normalizes them using TPM to generate expression matrices at the sample-cell subpopulation level.

[0046] The feature gene screening module incorporates the cross-cancer screening algorithm described in step 5 of Example 1, automatically completing 60% cancer type consistency screening and expression percentage secondary screening, and outputting a candidate gene set.

[0047] The core feature selection module integrates the LogisticRegressionCV function from the scikit-learn library, configures LASSO regularization parameters (the penalty coefficient C is optimized through 10-fold cross-validation), selects candidate gene sets, and outputs 39 core feature genes and their regression coefficients in each cell type.

[0048] The risk score calculation module uses the following formula: RiskScore i,c =∑ g∈Gc ( β g,c × X i,g,c ) The system automatically calculates the risk score for each sample across 11 immune cell subpopulations, generating an 11-dimensional feature vector.

[0049] The model training and evaluation module integrates machine learning libraries such as scikit-learn, XGBoost, and LightGBM, supporting automatic training and hyperparameter tuning of algorithms such as random forest, support vector machine, XGBoost, and neural networks (GridSearchCV). It includes built-in 5-fold cross-validation and independent test set evaluation functions, outputting performance metrics such as AUC, accuracy, sensitivity, and specificity.

[0050] The risk assessment module receives PBMC single-cell sequencing data from the subject, automatically calls the trained ensemble model, outputs the cancer risk probability value (0-100%) and risk level (high risk ≥80%, medium risk 30-80%, low risk <30%), and generates a visualization report showing the risk score contribution of each immune cell subset.

[0051] The system adopts a B / S architecture, with the front end developed using Vue.js, the back end using the FastAPI framework, and the database using MySQL to store sample information and prediction results. It supports batch sample processing and concurrent access by multiple users.

[0052] Example 3: Preparation and Application of a Multi-Cancer Risk Prediction Kit (1) Self-testing single-cell transcriptome sample collection and PBMC isolation To further test the stability of this model, we collected 28 clinical samples for single-cell transcriptome sequencing, including 23 cancer patients and 5 healthy controls. Fresh peripheral blood (2 mL) was collected from each subject and placed in an EDTA anticoagulant tube, diluted with an equal volume of 1×PBS (phosphate-buffered saline). PBMCs were then isolated using the Ficoll method, washed twice with 1×PBS, and counted to ensure that the cell count for each subject was ≥1×10⁻⁶. 6 indivual.

[0053] (2) Construction of single-cell sequencing libraries Each subject should take at least 1 x 10 6 One PBMC was used to load cell suspension onto a 10x Genomics Chromium single-cell sequencing platform (or other domestic platforms). GEMs (gel droplets) were generated using a microfluidic chip. Single-cell lysis and mRNA capture by barcode-bearing magnetic beads were completed in each droplet. Reverse transcription was performed in GEMs to convert mRNA into first-strand cDNA with cell tags and UMI (unique molecular identifier); reaction conditions: 53℃ for 45 minutes, 85℃ for 5 minutes; cDNA was recovered after demulsification and amplified by PCR (98℃ for 3 minutes; 98℃ for 15 seconds, 63℃ for 20 seconds, 72℃ for 1 minute, 12-14 cycles; 72℃ for 1 minute). Subsequently, SPRIselect magnetic beads were used for purification to remove excess primers and short fragments.

[0054] The amplified cDNA was fragmented, end-repaired, and A-tailed; an adapter (containing P5, P7 sequences and sample index) was ligated; the ligation product was enriched by final PCR amplification; the library was quality checked using an Agilent 2100 Bioanalyzer, requiring a library concentration ≥2nM and a main fragment peak between 400-600bp.

[0055] (3) Sequencing After the library is constructed and passes quality control, it is sequenced on the Illumina high-throughput sequencing platform, or the domestic BGI platform sequencer (DNBSEQ-T7) can be selected. After obtaining the Fastq data, Cell Ranger (version 7.0 or above) is used for data analysis. The quantified gene expression matrix data is input into the multi-cancer risk prediction system constructed in this invention and run according to the above steps to obtain the cancer detection rate of the 28 subjects.

[0056] See appendix Figure 3 The system outputs the test report for the subjects. In this batch of self-tested single-cell transcriptome data, the model successfully detected cancer with an accuracy rate of 96.1% (95% confidence interval: 0.946-0.976). These results indicate that this application demonstrates good detection performance in clinical samples and can be effectively used for early screening and risk assessment of multiple cancer types.

[0057] The above embodiments, in conjunction with the accompanying drawings, detail the entire process of this invention, from data construction and feature discovery to model training and validation. The results show that the ensemble learning model constructed based on PBMC single-cell atlases and 39 multi-cancer immune characteristic genes can efficiently and accurately distinguish between cancer patients and healthy individuals, demonstrating significant application potential in early risk screening for multiple cancer types.

[0058] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0059] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0060] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0061] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0062] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A method for constructing a multi-cancer risk prediction model based on PBMC single-cell sequencing, characterized in that, include: S1: Obtain single-cell sequencing data of PBMCs from various cancer patients and healthy controls to construct a pan-cancer PBMC single-cell atlas; S2, Based on the pan-cancer PBMC single-cell atlas, the data of each immune cell subset are integrated to construct a pseudo-batch expression matrix at the sample-cell subset level; S3, for each immune cell subset, the pseudo-batch expression matrix is ​​used to perform differential expression analysis between patient samples and healthy control samples of each cancer type to obtain the preliminary differentially expressed genes of each immune cell subset in each cancer type. S4, perform cross-cancer screening on the preliminarily differentially expressed genes to obtain a candidate gene set; S5. The candidate gene set is screened using the LASSO regression algorithm to obtain the expression data of the core characteristic gene set of each immune cell subset and its corresponding regression coefficient. S6, based on the expression data of the core feature gene set and its corresponding regression coefficients, calculates the risk score of each sample in each immune cell subpopulation to form a multidimensional risk feature vector. The multidimensional risk feature vector of the sample is used as input and trained using machine learning algorithms to obtain a multi-cancer risk prediction model.

2. The method for constructing a multi-cancer risk prediction model based on PBMC single-cell sequencing according to claim 1, characterized in that, In step S1, the various cancers include at least 15 types.

3. The method for constructing a multi-cancer risk prediction model based on PBMC single-cell sequencing according to claim 1, characterized in that, In step S2, the immune cell subsets include B cells and CD14 cells. + Monocytes, CD16 + Monocytes, CD4 + Conventional T cells, CD8 + There are 11 cell types, including T cells, dendritic cells, natural killer cells, plasma cells, platelets, proliferating T cells, and regulatory T cells.

4. The method for constructing a multi-cancer risk prediction model based on PBMC single-cell sequencing according to claim 1, characterized in that, In step S4, the cross-cancer screening includes: screening genes that show consistent expression trends in at least a preset proportion of cancer types, where consistent expression trends refer to both being upregulated or both being downregulated; and performing secondary screening based on the expression percentage of the screened genes in cancer types and healthy controls to obtain a candidate gene set, wherein the preset proportion is 60%.

5. The method for constructing a multi-cancer risk prediction model based on PBMC single-cell sequencing according to claim 1, characterized in that, In step S5, the core feature gene set contains 39 genes.

6. The method for constructing a multi-cancer risk prediction model based on PBMC single-cell sequencing according to claim 1, characterized in that, In step S6, the machine learning algorithm includes at least one of random forest, support vector machine, XGBoost, neural network, or a combination thereof.

7. A multi-cancer risk prediction system, characterized in that, Constructed using the method described in any one of claims 1 to 5, comprising: The data acquisition and atlas construction module is used to acquire single-cell sequencing data of PBMCs from various cancer patients and construct single-cell atlases of PBMCs from tumor patients. The pseudo-batch matrix construction module is used to integrate the data of each immune cell subset and construct a pseudo-batch expression matrix at the sample-cell subset level. The feature gene screening module is used to perform differential expression analysis of each immune cell subset across cancer types and obtain a candidate gene set; The core feature screening module is used to screen the candidate gene set using the LASSO regression algorithm to obtain the core feature gene set and its corresponding regression coefficients. The risk score calculation module is used to calculate the cell type-specific risk score for each sample based on the core feature gene set and its regression coefficients, forming a multidimensional risk feature vector. The model training and evaluation module is used to train and validate the multi-dimensional risk feature vectors through machine learning algorithms to obtain a multi-cancer risk prediction model. The risk assessment module is used to receive PBMC single-cell sequencing data from the subject and output the risk assessment results of the subject having multiple cancers based on the multi-cancer risk prediction model.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a method for constructing a multi-cancer risk prediction model based on PBMC single-cell sequencing as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements a method for constructing a multi-cancer risk prediction model based on PBMC single-cell sequencing as described in any one of claims 1 to 6.

10. Use of the multi-cancer risk prediction model constructed by the method of any one of claims 1 to 6 or the system of claim 7 in the preparation of a kit or diagnostic device for early cancer screening or risk assessment.

Images