Colorectal cancer liver metastasis detection method and system based on multi-omics deep learning

Abstract

The invention discloses a colorectal cancer liver metastasis detection method and system based on multi-omics deep learning, and the method comprises the steps: extracting a characteristic gene group with CRC metastasis-metabolism dual functions, and on the basis of constructing a CRC metastasis risk prediction model, screening molecular markers with CRC metastasis general and CRLM specific characteristics based on the analysis of CRC metastasis risk probability, the biomarker is used as an initial candidate biomarker; meanwhile, a CRLM patient sample and differential expression metabolites corresponding to an LCRC patient sample metabolome are obtained from a metabolome, and a CRLM multilayer core metabolic network is constructed in combination with enrichment pathway analysis of a transcriptome and the metabolome, so that core biomarkers are screened for accurate detection of colorectal cancer liver metastasis; the whole method can effectively screen the biomarker beneficial to detection of colorectal cancer liver metastasis.

Description

Technical Field

[0001] This invention belongs to the technical field of combining bioinformatics, deep learning and tumor molecular diagnostics, and specifically relates to a method and system for detecting colorectal cancer liver metastases based on multi-omics deep learning. Background Technology

[0002] Colorectal cancer (CRC) is the third most common malignant tumor worldwide, and liver metastasis (CRLM) is one of the leading causes of death. Early detection and assessment of liver metastasis risk can significantly improve resectable rates and prognosis.

[0003] Current clinical early CRLM screening mainly relies on imaging (CT, MRI, PET, ultrasound) combined with serum tumor markers (such as CEA, CA19-9). These methods have high sensitivity in detecting established metastases, but their predictive ability is limited before surgery and in the early stages of tumor development. The spatial resolution and detection window of imaging often make it difficult to identify small metastases before they develop into larger lesions. These early signs of lesions are often detected through molecular testing, while conventional serum markers have insufficient specificity in patients with early liver metastases, leading to a high rate of false positives or false negatives.

[0004] In recent years, some studies have utilized genomic or transcriptomic data for prognostic assessment, such as the status of driver genes like KRAS / NRAS and BRAF, or microsatellite stability (MSS) and instability (MSI) typing. These methods can guide targeted or immunotherapy regimens for patients with liver metastases to some extent, but they still have significant limitations in early screening for liver metastases: limited information dimensions: a single genome cannot capture the metabolic reprogramming of tumors and the complex changes in the immune microenvironment; insufficient predictive performance: because the metastasis process is regulated by multiple factors, including metabolic state, immune escape mechanisms, and microenvironmental heterogeneity, relying solely on genotype or local transcriptomic features is insufficient to achieve high sensitivity and specificity in prediction, and is also limited to explaining the biological mechanisms of metastasis; poor cross-platform portability: significant batch effects exist between different sequencing platforms and cohorts, and the generalization performance of a single sequencing platform on different datasets is low.

[0005] Some studies have attempted to integrate transcriptomic, proteomic, and metabolomic data for tumor subtyping or mechanistic research. For example, Sinkala et al. analyzed the association between metabolic gene variations and drug responses in various cancers. However, most of these multi-omics methods remain at the level of statistical correlation or differential analysis, based on relatively direct linear relationships and simple feature dimension analysis. They lack analysis of complex high-dimensional features and nonlinear relationships in real medical scenarios and lack high-performance predictive models for early clinical screening. Furthermore, in practical applications, multi-omics models often encounter the following difficulties: Cross-platform data integration is difficult: data sources are diverse and batch differences are significant, direct integration easily introduces noise and weakens the model's generalization ability; High-dimensional data modeling limitations: metabolomics and transcriptomics feature dimensions are large, with many nonlinear relationships. Traditional machine learning models (such as random forests and support vector machines) are insufficient in analyzing nonlinear relationships under high-dimensional data, prone to overfitting and exhibiting poor generalization ability; Lack of precise biomarkers: existing biomarker screening in multi-omics studies often lacks mechanism validation and clinical application evaluation, making it difficult to implement in specific scenarios such as preoperative prediction, risk stratification, and treatment decision support.

[0006] In summary, existing liver metastasis risk prediction technologies share the following common shortcomings: poor early prediction performance: imaging and traditional biomarkers have limited predictive ability in patients without significant tumor burden in the early stages; incomplete utilization of omics information: single omics studies ignore key information such as metabolic characteristics and immune environment, reducing discriminative power and the ability to analyze comprehensive biological functions; poor cross-cohort stability: insufficient model adaptability and portability, and a lack of robust batch calibration framework; lack of clinical validation of biomarkers: it is difficult to provide credible and validated molecular targets while assessing risk.

[0007] Therefore, there is an urgent need for a deep learning framework that can operate stably across platforms and integrate multi-omics data, enabling early prediction and identification of biomarkers with mechanistic significance and clinical translational potential, thereby supporting the development of personalized monitoring and treatment strategies. Summary of the Invention

[0008] In view of the above, the purpose of this invention is to provide a method and system for detecting colorectal cancer liver metastases based on multi-omics deep learning. This method combines large-scale transcriptomic analysis, deep learning modeling, and a cross-omics approach using serum metabolomics to facilitate early detection, biomarker discovery, risk assessment, and personalized treatment strategy support for colorectal cancer liver metastases (CRLM). This method and system can be extended to predicting metastasis risk and studying the molecular mechanisms of other malignant tumor types.

[0009] To achieve the above-mentioned objectives, an embodiment provides a method for detecting colorectal cancer liver metastases based on multi-omics deep learning, comprising the following steps: Feature gene clusters with dual functions of CRC transfer-metabolic reprogramming were screened based on transcriptome datasets, and batch effect correction was performed on large-scale transcriptomics cohorts for deep learning and prediction. Using characteristic gene clusters as input, a liver metastasis risk prediction model was constructed by deep learning of CNN based on the characteristic gene clusters. The transcriptome risk score was calculated based on the metastasis risk probability of the CRC metastasis risk prediction model, and the correlation analysis between metastatic CRC (hereinafter referred to as mCRC) and CRLM occurrence results was performed based on the transcriptome risk score. Highly correlated features with CRC metastasis commonality and CRLM specificity were screened as initial candidate biomarkers. Differentially expressed metabolites corresponding to the metabolomes of CRLM and LCRC patient samples were obtained. Enrichment pathway analysis was performed on the initial candidate biomarkers corresponding to the transcriptome and the differentially expressed metabolites corresponding to the metabolome. A CRLM multilayer core metabolic network with shared enriched pathways was constructed, and candidate biomarkers within the pathways were labeled. The CRLM multilayer core metabolic network was used to test the CRLM resolution performance of the candidate biomarkers, and core biomarkers were screened. The core biomarkers are used for the precise detection of colorectal cancer liver metastases.

[0010] Preferably, batch effect correction based on transcriptome dataset includes: All transcriptome datasets were merged according to common genes, and the sequencing platform was used as a batch variable and disease status as a biological grouping variable. ComBat was used for batch effect correction, while retaining the biological information related to disease status. The correction effect was visualized and verified using PCA, and the batch effect was confirmed to be reduced when the mean silhouette coefficient of the samples was ≤ 0.1.

[0011] Preferably, a transcriptome risk score is calculated based on the liver metastasis risk probability in the liver metastasis risk prediction model using characteristic gene clusters from patient samples, including: The average probability of liver metastasis risk in the five sub-models included in the liver metastasis risk prediction model for the characteristic gene groups of the patient sample was calculated as the transcriptome risk score.

[0012] Preferably, the association between metastatic CRC and CRLM is analyzed based on transcriptome risk scores to screen for highly correlated features with both CRC metastasis universality and CRLM specificity as initial candidate biomarkers, including: High-transfer risk features were screened from the characteristic gene population based on transcriptome risk scores. The correlation coefficients between all features and transcriptome risk scores in the mCRC and CRLM cohorts were calculated. The N features ranked by correlation coefficients in the mCRC and CRLM cohorts were selected and their intersection was calculated to obtain the core features. These core features are molecular markers with CRC transfer generality and CRLM specificity, and are used as initial candidate biomarkers.

[0013] Preferably, the shared enrichment pathway refers to the key metabolic pathways that are simultaneously enriched by cross-analysis of the metabolic pathways enriched by the initial candidate biomarkers and the metabolic pathways enriched by the differentially expressed metabolites, including the retinol metabolic pathway and the tryptophan metabolic pathway.

[0014] Preferably, the CRLM multilayer core metabolic network with shared enrichment pathways includes key metabolic pathways, biomarkers of regulatory genes, transcription factors, and metabolites, wherein transcription factors act on biomarkers, biomarkers act on pathways, and pathways produce metabolites. Based on the CRLM multilayer core metabolic network, a CRLM core metabolic feature map was constructed, and candidate biomarkers were marked in the CRLM core metabolic feature map.

[0015] Preferably, the CRLM multilayer core metabolic network is used to test the CRLM resolution performance of candidate biomarkers and screen core biomarkers, including: (1) Performance evaluation of candidate biomarkers: All candidate biomarkers labeled in the CRLM multilayer core metabolic network were used to distinguish between colorectal cancer liver metastases (CRLM) samples and colorectal cancer primary lesions (CRC PR) samples. The discrimination performance of each candidate biomarker was quantified by calculating the area under the receiver operating characteristic curve of each candidate biomarker, and the biomarker with the best discrimination performance and the highest sensitivity was selected as the candidate biomarker for early screening of CRLM. (2) Performance verification and determination of core biomarkers: The optimal candidate biomarkers selected in step (1) are compared with multiple reported CRLM predictive biomarkers to evaluate the sensitivity of the candidate biomarkers to CRLM patients. The difference in sensitivity of the candidate biomarkers to CRLM patients compared with known CRLM screening biomarkers is evaluated by comparing ROC curves and AUC values. The core biomarkers are determined based on the evaluation results.

[0016] Preferably, the method further includes: verifying the mechanism of action of core biomarkers in key metabolic pathways by combining metabolic monitoring data, and realizing metabolic regulation based on the mechanism of action, i.e., performing metabolic reprogramming.

[0017] To achieve the above-mentioned objectives, embodiments of the present invention also provide a colorectal cancer liver metastasis detection system based on multi-omics deep learning, comprising: A dual-function feature screening module is used to screen feature gene groups with dual functions of CRC transfer-metabolic reprogramming based on transcriptome datasets, and to perform sample batch effect correction on transcriptomics cohorts for deep learning and prediction. The initial candidate biomarker screening is used to construct a liver metastasis risk prediction model by using a characteristic gene group as input and deep learning of a CNN using the characteristic gene group. The transcriptome risk score is calculated based on the metastasis risk probability of the CRC metastasis risk prediction model, and the correlation analysis between metastatic CRC and CRLM is performed based on the transcriptome risk score. Highly correlated features with CRC metastasis commonality and CRLM specificity are screened as initial candidate biomarkers. The core biomarker screening module is used to obtain differentially expressed metabolites corresponding to the metabolomes of CRLM and LCRC patient samples. It performs enrichment pathway analysis on the initial candidate biomarkers corresponding to the transcriptome and the differentially expressed metabolites corresponding to the metabolome, constructs a CRLM multilayer core metabolic network with shared enriched pathways, and labels candidate biomarkers within the pathways. The CRLM multilayer core metabolic network is used to test the CRLM resolution performance of candidate biomarkers, and screens core biomarkers for precise detection of colorectal cancer liver metastases.

[0018] Compared with the prior art, the beneficial effects of the present invention include at least the following: This invention extracts characteristic gene clusters with dual functions of CRC metastasis and metabolism, and constructs a CRC metastasis risk prediction model. Based on the probability of CRC metastasis risk, it screens molecular biomarkers with CRC metastasis commonality and CRLM specificity as initial candidate biomarkers. At the same time, from the perspective of metabolomics, it obtains differentially expressed metabolites corresponding to the metabolomes of CRLM patient samples and LCRC patient samples, and constructs a multi-layer core metabolic network of CRLM by combining transcriptome and metabolome enrichment pathways. Based on this, it is used to screen core biomarkers for accurate detection of colorectal cancer liver metastasis. The whole method can effectively screen biomarkers that are beneficial for detecting colorectal cancer liver metastasis. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the colorectal cancer liver metastasis detection method based on multi-omics deep learning provided in the embodiment; Figure 2 This is the feature candidate pool construction process provided in the embodiment; Figure 3 This is the CRC transfer risk prediction model and transcriptome risk score calculation process provided in the embodiment; Figure 4 The examples show the screening results and batch effect correction effects of the characteristic gene groups provided. Figure 5 This is a schematic diagram of the structure of a one-dimensional convolutional neural network provided in the embodiment; Figure 6 This is a graph showing the training process, internal validation, and external validation performance of the CRC transfer risk prediction model provided in the embodiment. Figure 7 This is a benchmark test graph showing the performance of the model of the present invention provided in the embodiments and 10 classic machine learning classifiers in identifying patients with metastatic colorectal cancer. Figure 8 and Figure 9 This is a screening chart of features related to transcriptome risk score (TRS) provided in the examples; Figure 10 This is a functional analysis diagram of features related to transcriptome risk score (TRS) provided in the examples; Figure 11 This is the CRLM multilayer core metabolic network construction process provided in the embodiment; Figure 12 and Figure 13 This is an intersection analysis of the initial candidate biomarkers (genes that promote metastasis function, i.e., TRS core features) provided in the examples and the metabolite enrichment pathways that regulate metastasis function, in order to identify key CRLM-related pathways. Figure 14 This is a CRLM multilayer core metabolic network diagram provided in the embodiments; Figure 15 This is the method provided in the embodiments for selecting the optimal marker from candidate markers; Figure 16 The CRC cell line provided in the examples verifies the ACMSD-induced metastasis phenotype; Figure 17 This is a clinical validation of the metastatic phenotype of the biomarker ACMSD provided in the examples; Figure 18 The CRC cell line provided in the examples is used to verify the mechanism by which ACMSD promotes metastasis. Figure 19 This is a structural diagram of the colorectal cancer liver metastasis detection system based on multi-omics deep learning provided in the embodiment. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0022] The inventive concept of this invention addresses the following problems existing in the early detection and risk assessment technologies for colorectal cancer liver metastases (CRLM): Current methods relying on imaging and conventional serum tumor markers have insufficient sensitivity and specificity in predicting liver metastasis risk in the early clinical stage, especially when metastatic lesions have not yet formed significant imaging features; molecular detection methods based on single omics (such as genomics or transcriptomics) have limited information dimensions, neglecting important mechanistic factors such as metabolic reprogramming and changes in the immune microenvironment, and cannot accurately characterize the multifactorial effects in the process of liver metastasis; existing models have poor generalization ability across sequencing platforms or different patient cohorts, and lack batch effect correction and feature extraction frameworks that can run stably in multi-source cohorts; existing omics studies lack mechanism validation and clinical application evaluation in marker screening, making it difficult to directly translate results into accurate preoperative prediction, risk stratification, and medication guidance.

[0023] To address the above issues, the technical solution proposed in this invention combines large-scale transcriptomics analysis, deep learning models, and a cross-omics integration method of serum metabolomics to establish an analytical framework that can stably operate on multiple data sources, achieving the following technical tasks: When liver metastases are not yet visible on clinical imaging, the risk of liver metastasis in individual patients can be predicted with high accuracy through gene and molecular characteristics, enabling early screening of CRC patients for metastasis, early detection of patients with high metastasis risk for medical intervention, and saving personal and public medical costs; Based on cross-omics characteristics, metabolic pathways and gene molecules closely related to liver metastasis can be identified and pinpointed, discovering novel biomarkers with mechanistic significance; Through clinical samples and in vitro validation, the association between biomarkers and pathological stage, recurrence risk, immune microenvironment characteristics, and sensitivity to specific targeted drugs can be clarified, providing quantifiable evidence for personalized treatment strategies and improving the effectiveness of precision treatment for patients; Therefore, this invention aims to effectively overcome the shortcomings of existing early prediction technologies for colorectal cancer liver metastasis in terms of information utilization, model stability, biomarker screening and validation, and to achieve the integrated technical goal of metastasis risk assessment, mechanism analysis and treatment decision support.

[0024] like Figure 1 As shown in the embodiment, the method for detecting colorectal cancer liver metastases based on multi-omics deep learning includes the following steps: S1, based on the transcriptome dataset, screened characteristic gene groups with dual functions of CRC transfer-metabolic reprogramming, and performed sample batch effect correction on the transcriptomics cohort.

[0025] In this embodiment, transcriptome datasets containing clinical annotations for colorectal cancer were obtained from multiple public databases (such as GEO, cBioPortal, TCGA, GSE131418, etc.). The data types included microarray expression profiling (GPL570 platform) and RNA-seq expression data (Illumina NGS sequencing). The clinical labels of these data were distinguished as primary CRC (CRC-PR) and metastatic CRC (mCRC), and pathological M stage was marked in samples involved in CRLM.

[0026] In the embodiments, such as Figure 2 As shown, when screening for feature gene clusters with dual functions of CRC metastasis and metabolism, cancer metabolism-related genes were extracted from the published collection of tumor metabolic reprogramming genes. The intersection of these genes with differentially expressed genes (DEGs) related to mCRC vs CRC-PR regulation in the GSE131418 dataset (which contains the largest CRC sample) was obtained to obtain feature gene clusters related to CRC metastasis regulation and metabolic function regulation. These feature gene pools with dual functions of CRC metastasis and metabolism were then used as feature candidate pools for model input.

[0027] In the embodiments, such as Figure 3 As shown, during cross-platform batch effect correction, all transcriptome datasets were merged according to common genes, and sequencing platform was used as batch variable and disease state as biological grouping variable. ComBat (svaR package) was used for batch effect correction, while retaining biological information related to disease state. The correction effect was verified by PCA visualization, and the mean silhouette coefficient of the samples was calculated to be ≤ 0.1 to confirm that the batch effect was reduced and the sample distribution changed from platform clustering to cross-platform mixing, ensuring that the model can run stably in different data sources and maintain predictive performance.

[0028] In the examples, the screening results of the characteristic gene population and the batch effect correction effect are specifically given, such as... Figure 4 As shown. Figure 4 In the middle A, the results of principal component analysis (PCA) of the metastatic colorectal cancer (mCRC) and primary colorectal cancer (CRC PR) groups in the GSE131418 dataset are shown. Figure 4 In Figure B, a volcano plot shows the differentially expressed genes (DEGs) between mCRC and CRC PR. Genes with a |log2 fold change| > 1 and P < 0.05 are considered DEGs. A total of 4,067 mCRC-related DEGs were identified from GSE131418. Figure 4 The Venn diagram shown in C illustrates the intersection between DEGs and a predefined set of metabolic genes in 32 human cancers, resulting in 620 functional DEGs. Figure 4 China D and Figure 4 E in the figure represents the cross-platform PCA diagrams before and after ComBat batch effect correction, and after correction. Figure 4 F represents the mean silhouette score coefficients before and after batch effect correction to assess clustering trend. A lower silhouette score coefficient after correction indicates that the batch effect has been effectively reduced. Statistical differences between groups were assessed using t-tests. The number of "*" symbols indicates significance: *P < 0.05; **P < 0.01; ***P < 0.001; ns, not significant.

[0029] S2. Using the characteristic gene cluster as input, a liver metastasis risk prediction model is constructed by deep learning of CNN using the characteristic gene cluster. The transcriptome risk score is calculated based on the metastasis risk probability of the CRC metastasis risk prediction model. The correlation between metastatic CRC and CRLM is analyzed based on the transcriptome risk score. Highly correlated features with CRC metastasis commonality and CRLM specificity are screened as initial candidate biomarkers. In the embodiments, such as Figure 3 As shown, the batch-corrected patient samples are divided into training and validation sets proportionally; within each cross-validation layer, the Z-score standardized parameters are calculated and applied to the validation and test sets.

[0030] like Figure 5 As shown, the model architecture of the CRC transfer risk prediction model adopts a one-dimensional convolutional neural network (1D-CNN), which is used to extract transcriptomic features from the feature gene group and predict the probability of CRC transfer risk. The specific structure is as follows: First convolutional-pooling module: Convolutional layer (Conv1D) (number of kernels = 32, kernel size = 5, activation function ReLU) + Max pooling layer (MaxPooling1D) (pooling window size = 2 + dropout rate = 0.3); ReLU activation function formula: f(x) = max(0,x); Second convolutional-pooling module: Convolutional layer (Conv1D) (number of kernels = 64, kernel size = 5, activation function ReLU) + Max pooling layer (MaxPooling1D) (pooling window size = 2 + dropout rate = 0.4); Fully connected module: Flatten layer + Dense layer (64 neurons, activation function ReLU, dropout rate = 0.5); Output layer: Dense layer (1 neuron, activation function Sigmoid) with output transition risk probabilities of 0 and 1.

[0031] In this embodiment, when training the CNN using the feature gene clusters from the feature candidate pool, the loss function is binary cross-entropy, the optimizer is Adam, and the initial learning rate is 0.001. Five-fold hierarchical cross-validation is also performed (4 folds for training, 1 fold for validation), with each fold generating a sub-model, for a total of 5 sub-models. Each sub-model uses an early stopping strategy (patience = 10) and a gradual learning rate decay strategy (Reduce LR On Plateau, decay factor = 0.5, patience = 5) to avoid overfitting. For CRC metastasis risk prediction, the feature gene clusters of patient samples are input into the five sub-models for ensemble training and prediction, and the metastasis risk probability is calculated after forming the ensemble model.

[0032] The embodiments also provide the training process, internal verification, and external verification performance graphs of the CRC transfer risk prediction model of the present invention, such as... Figure 6 As shown, Figure 6 In the figure, A represents the loss curves of each sub-model over time on the training and validation sets, used to evaluate training convergence and potential overfitting. Figure 6 Accuracy curves of each sub-model in B on the training and validation sets. Figure 6 C represents the area under the receiver operating characteristic (AUC) curve on the validation set of each fold model. Figure 6 In the figure, D represents the precision and recall metrics of each sub-model on the validation set. Figure 6 E in the diagram represents the confusion matrix of the internal validation set. Figure 6 China F and Figure 6 G represents the receiver operating characteristic (ROC) curve and precision-recall (PR) curve of the internal validation set, along with the corresponding AUC value. Figure 6 H in the diagram represents the confusion matrix of the external test queue (CRLM). Figure 6 I and Figure 6 J represents the Receiver Operating Characteristic (ROC) curve and precision-recall (PR) curve for the external test set, along with the corresponding AUC value. In internal five-fold cross-validation, the CRC transfer risk prediction model of this invention achieved an AUC of 0.92 for mCRC vs CRC-PR classification, with a precision of 86% and an accuracy of 85%. On the external independent test set (CRLM dataset), the model performance remained stable, achieving an AUC of 0.97 for CRLM vs CRC-PR classification, with an accuracy of 92% and an accuracy of 94%.

[0033] The embodiments also provide a comparison between the model of the present invention and 10 high-performance machine learning models such as LightGBM, Random Forest, and XGBoost. Figure 7 As shown, Figure 7 In the diagram, A represents the classification performance comparison on the internal validation set, where the x-axis represents AUC and the y-axis represents accuracy. The model located in the upper right quadrant represents the best performance. Figure 7 Table B represents the classification performance comparison on the external test set, where the x-axis represents AUC and the y-axis represents accuracy. Models located in the upper right quadrant represent the best performance. Tables 1 and 2 also illustrate the accuracy, AUC, precision, and recall of each model on the internal validation set and the external test set. Analysis Figure 7 As shown in Tables 1 and 2, the method of the present invention exhibits the smallest performance degradation on the independent test set, the best classification performance, and significantly better generalization ability than existing methods.

[0034] Table 1. Baseline test results of classification performance on the validation set.

[0035] Table 2. Baseline Classification Performance Tests on External Test Sets

[0036] like Figure 3 As shown, when calculating the transcriptomic risk score based on the CRC metastasis risk probability, the average of the liver metastasis risk probabilities of the characteristic gene groups of the patient sample in the five sub-models included in the CRC metastasis risk prediction model is used as the transcriptomic risk score (TRS).

[0037] In this embodiment, the correlation between metastatic CRC and CRLM was also analyzed based on the transcriptome risk score (TRS). Highly correlated features with both CRC metastasis commonality and CRLM specificity were screened as initial candidate biomarkers. The specific process is as follows: based on the transcriptome risk score (TRS), high metastasis risk features were screened from the characteristic gene population. A threshold of 0.5 was set. When TRS ≥ 0.5, the characteristic gene population of the corresponding patient was considered high risk, that is, the probability of metastasis was higher. When TRS < 0.5, it was considered low risk, that is, the probability of metastasis was lower.

[0038] During feature association analysis, the correlation coefficients between all features in the mCRC cohort and the CRLM cohort and the transcriptome risk score TRS were calculated. The top N (e.g., 50) features with the highest correlation coefficients in the mCRC and CRLM cohorts were selected and their intersection was calculated to obtain the core features. These core features are molecular markers with CRC transfer commonality and CRLM specificity, and are used as initial candidate biomarkers for early CRLM screening.

[0039] The examples specifically provide screening and functional analysis plots of features related to transcriptome risk score (TRS). Figure 8 To display the top 20 features with the strongest positive or negative association with TRS in the merged mCRC and CRC PR queues (training set and internal validation set). Figure 9 To display the top 20 features with the strongest positive or negative correlation to TRS in external test queues (CRLM and CRC PR), where correlation is used. Figure 10 A in the middle is a pair Figure 8 and Figure 9 By using Spearman correlation to evaluate the intersection of the top 50 TRS-related features identified in the assessment, 22 core TRS features were obtained. Figure 10 In the B-group, pathway enrichment analysis was performed on 22 core TRS features, highlighting significantly enriched metabolic and signaling pathways. Figure 10 In the middle (C), the difference in core TRS features between CRLM and CRC PR samples in the GSE50760 dataset is presented as log2(FPKM+1). Red boxes represent CRLM samples, and gray boxes represent CRC PR samples; significance levels are marked with an asterisk. Statistical differences between groups were assessed using t-tests. *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001; ns, not significant.

[0040] S3. Differentially expressed metabolites corresponding to the metabolomes of CRLM and LCRC patient samples were obtained. Enrichment pathway analysis was performed on initial candidate biomarkers corresponding to the transcriptome and differentially expressed metabolites corresponding to the metabolome. A CRLM multilayer core metabolic network with shared enriched pathways was constructed, and candidate biomarkers within the pathways were labeled. The CRLM multilayer core metabolic network was used to test the CRLM resolution performance of the candidate biomarkers, and core biomarkers were screened. These core biomarkers are used for precise detection of colorectal cancer liver metastases, such as... Figure 11 As shown.

[0041] In this embodiment, differentially expressed metabolites were also obtained from the metabolomics data. The analysis process for the metabolomics data followed conventional existing techniques. Specifically, serum samples from CRLM patients and orthotopic CRC (LCRC) patients were first subjected to standard LC-MS / MS non-targeted metabolomics detection using a Waters HSS T3 C18 column and gradient elution. ESI source and MRM detection were performed according to standard parameters. Then, differential metabolite analysis was performed. Specifically, orthogonal partial least squares discriminant analysis (OPLS-DA) was used to distinguish the component differences between metabolomics sample groups. Variable importance inprojection analysis was used to screen for differentially expressed metabolites (DEMs) with VIP>1 and P<0.05.

[0042] In this embodiment, enrichment pathway analysis was also performed on the initial candidate biomarkers corresponding to the transcriptome. Specifically, the initial candidate biomarkers were enriched in the GO and KEGG biopathology databases, showing significant enrichment in liver metastasis-related metabolic pathways such as retinol metabolism and tryptophan metabolism. Pathway enrichment analysis was also performed on differentially expressed metabolites detected by metabolomics. Cross-analysis was then conducted on the pathways enriched by the initial candidate biomarkers and the pathways enriched by the differentially expressed metabolites; the intersection was used to identify the key pathways co-enriched across omics, which were designated as common enriched pathways. These included the two key pathways driving CRLM: retinol metabolism and tryptophan metabolism (GSEA NES > 2.2, P.adjust < 0.001).

[0043] Figure 12 and Figure 13 We present an intersection analysis of initial candidate biomarkers (genes that promote metastasis, i.e., TRS core features) and enriched pathways of metabolites that regulate metastasis to identify key CRLM-related pathways. Figure 12 The metabolomics analysis workflow was as follows: serum samples were collected from patients with in situ colorectal cancer (LCRC, n = 15) and colorectal cancer liver metastases (CRLM, n = 15), and metabolites were sequenced by chromatography-mass spectrometry (LC-MS), followed by metabolite clustering, differential expression analysis, and pathway enrichment analysis. Figure 13 In the middle, A is the score plot of Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA), showing a clear separation between CRLM and LCRC samples. Figure 13 In the volcano plot (B), differentially expressed metabolites (DEMs) associated with CRLM are highlighted. Metabolites were identified as DEMs based on a |log2 fold change| > 1 or < −1 and a P-value < 0.05. Figure 13The Venn diagram in the middle illustrates the overlap between pathways enriched from the initial candidate biomarkers and those enriched from DEMs. Only pathways satisfying p.adjust < 0.05 or FDR significance are retained for intersection analysis. Figure 13 The diagram in section D summarizes the cross-pathways and their significant enrichment in transcriptomics and metabolomics analyses. Retinol metabolism and tryptophan metabolism are among the most representative enriched pathways. Figure 13 China E and Figure 13 In the figure, F represents the gene set enrichment analysis (GSEA) curves for retinol metabolism pathway and tryptophan metabolism in the CRLM sample (GSE50760 human liver metastasis sequencing dataset).

[0044] In the embodiments, a CRLM multilayer core metabolic network with shared enrichment pathways was also constructed, such as Figure 14 As shown, candidate biomarkers within the pathway are labeled. Specifically, the CRLM multilayer core metabolic network, which enriches the pathway, includes key metabolic pathways, biomarkers of regulatory genes, transcription factors, and metabolites. Transcription factors act on biomarkers, biomarkers act on pathways, and pathways produce metabolites. Based on the CRLM multilayer core metabolic network, a CRLM core metabolic feature map is drawn, and candidate biomarkers are labeled in the CRLM core metabolic feature map.

[0045] In this embodiment, the CRLM multilayer core metabolic network was also used to test the CRLM resolution performance of candidate biomarkers, and core biomarkers (such as ACMSD) were screened. These core biomarkers are used for the precise detection of colorectal cancer liver metastases. The specific process is as follows: (1) Performance evaluation of candidate biomarkers: All candidate biomarkers labeled in the CRLM multilayer core metabolic network (including ACMSD, ADH1A, CYP2C8, CYP1A2, RDH16, ALDH8A1, CYP2A6, CYP2C9, HSD17B6, and DHTKD1) were used to distinguish between colorectal cancer liver metastases (CRLM) samples and primary colorectal cancer lesions (CRC PR) samples. The discriminative performance of each candidate biomarker was quantified by calculating the area under the receiver operating characteristic (ROC) curve (AUC). The biomarker with the best discriminative performance and highest sensitivity was selected as the candidate biomarker for early CRLM screening. Figure 15 A.

[0046] The AUC value of the ACMSD was 0.97 (95% confidence interval of 0.93-1), demonstrating the best discriminative performance among all candidate biomarkers.

[0047] (2) Performance verification and determination of core biomarkers: The optimal candidate biomarkers selected in step (1) were compared with several reported CRLM predictive biomarkers (including CXCL12, FGF19, CXCR4, EGFR, CDX2, ERBB3, ERBB2, and PTGS2) to evaluate the sensitivity of the candidate biomarkers in distinguishing CRLM patients. ROC curve and AUC value comparison analyses were used to evaluate the sensitivity differences of the candidate biomarkers compared with known CRLM screening biomarkers for CRLM patients. Based on the evaluation results, the core biomarker was determined. Among them, the candidate biomarker ACMSD showed superior classification accuracy in distinguishing CRLM from CRC samples compared to all reported biomarkers; ultimately, ACMSD was determined as the core biomarker for accurate detection of colorectal cancer liver metastases. Figure 15 B.

[0048] Most biomarkers in the metabolic network have strong discriminative power for CRLM patients (AUC: 0.89-0.97), and ACMSD is the molecular marker with the best discriminative power for CRLM patients. It is also a novel molecular marker for CRLM screening. Compared with the molecular markers currently reported for the diagnosis of CRLM patients (ERBB2, FGF19 and CDX2, etc.), ACMSD is still the molecular marker with the best discriminative power for CRLM patients.

[0049] In this embodiment, metabolic monitoring data were used to verify the mechanism of action of core biomarkers in key metabolic pathways, and metabolic regulation was achieved based on the mechanism of action, i.e., metabolic reprogramming. Specifically, for ACMSD, the role of ACMSD in tryptophan-NAD was verified. + Mechanism of action in metabolic pathways (promotes the conversion of ACMS to pyridoxine, reduces quinic acid and NAD) + (Synthetic efficiency) and leads to metabolic reprogramming.

[0050] In the embodiments, the core biomarkers selected were also clinically validated and molecularly characterized to verify their effectiveness in detecting colorectal cancer liver metastases. Specifically, immunohistochemical validation (IHC), immune microenvironment analysis, and drug sensitivity prediction and drug treatment strategies were performed.

[0051] For immunohistochemical validation (IHC), ACMSD antibody was used to detect tumor tissue sections from patients with CRLM and LCRC, quantifying expression levels. Using the CRC cell lines SW480 and NCI-H716, SW480 was knocked down using biotechnology (shRNA lentiviral infection) to construct a CRC cell line with the target gene knocked down. Figure 16 China A and Figure 16As shown in Figure B, Western blotting (WB) experiments confirmed that ACMSD protein expression was knocked down after SW480 cells were infected with shRNA virus. Transwell assays were used to verify the regulatory effect of the target gene ACMSD on CRC cell metastasis, and the data on metastatic cells were quantified. Figure 16 C and Figure 16 Middle D. Tissue sections reflect the protein expression levels of ACMSD between CRLM and LCRC patient tissues, such as Figure 17 China A; Figure 17 The B-means index reflects the protein expression level of ACMSD in the tissues of patients with CRLM and LCRC. It also analyzes the statistical association between ACMSD expression and the risk of postoperative recurrence in CRC patients, such as... Figure 17 C; Analysis of ACMSD expression in patient tissues and its correlation with the pathological T-Stage stage of CRC patients, such as Figure 17 In terms of molecular mechanisms, after the target gene ACMSD is knocked down (e.g., ... Figure 18 In the middle A pathway, both the EMT (epithelial-mesenchymal transition) pathway, which promotes cell metastasis, and the TGF-β signaling pathway, which promotes cell metastasis, were significantly inhibited. Figure 18 B- Figure 18 As shown in Figure E, the results of cell validation demonstrate that the biomarkers discovered by this workflow exhibit high sensitivity for metastasis in CRC patients, and possess observable-verifiable and molecularly interpretable identification capabilities, thereby validating the usability and feasibility of the workflow.

[0052] For the analysis of the immune microenvironment, the ssGSEA algorithm was used to calculate the infiltration of immune cells in CRLM patient tissues and to evaluate the differences in immune activation and immunosuppressive cells in the ACMSD high expression group. Combined with the sequencing data of CRLM patient tissues, the expression analysis of immune checkpoint genes (PDCD1, LAG3, HAVCR2, etc.) was used to evaluate the potential for immunotherapy.

[0053] For drug sensitivity prediction and drug treatment strategies, the correlation between ACMSD expression and the sensitivity of CRLM patients to treatments (chemotherapy and targeted therapy) was calculated based on real-world data (a dataset of cohorts of CRLM patients with administration and administration outcomes); the correlation between ACMSD expression and the sensitivity of CRC targeted therapy (EGFR / VEGFR inhibitors) was calculated based on administration data from GDSC (Genomics of Drug Sensitivity in Cancer, a cellular drug sensitivity database); and the OncoPredict algorithm was used to model and predict the probability of benefit for patients with high ACMSD expression under specific targeted therapy regimens.

[0054] like Figure 19As shown in the embodiment, an embodiment also provides a colorectal cancer liver metastasis detection system 30 based on multi-omics deep learning, including: a dual-function feature screening module 31, an initial candidate biomarker screening module 32, and a core biomarker screening module 33. The dual-function feature screening module 31 is used to perform batch effect correction on samples based on a transcriptome dataset and screen feature gene clusters with dual CRC metastasis-metabolic functions. The initial candidate biomarker screening module 32 is used to construct a liver metastasis risk prediction model using deep learning on a CNN with the feature gene clusters, calculate a transcriptome risk score based on the liver metastasis risk probability of the patient sample's feature gene clusters in the liver metastasis risk prediction model, and perform CRC feature and... Feature association analysis of CRLM characteristics is used to screen molecular biomarkers with CRC metastasis commonality and CRLM specificity as initial candidate biomarkers. The core biomarker screening module 33 is used to obtain differentially expressed metabolites corresponding to the metabolomes of CRLM patient samples and LCRC patient samples. Enrichment pathway analysis is performed on the initial candidate biomarkers corresponding to the transcriptome and the differentially expressed metabolites corresponding to the metabolome. A CRLM multilayer core metabolic network with common enriched pathways is constructed, and candidate biomarkers within the pathway are labeled. The CRLM resolution performance of the candidate biomarkers is tested using the CRLM multilayer core metabolic network to screen core biomarkers. The core biomarkers are used for accurate detection of colorectal cancer liver metastasis.

[0055] It should be noted that the colorectal cancer liver metastasis detection system based on multi-omics deep learning provided in the above embodiments should be illustrated using the above-described functional module division as an example when performing accurate detection of colorectal cancer liver metastases. The functions described above can be assigned to different functional modules as needed, i.e., the internal structure of the terminal or server can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the colorectal cancer liver metastasis detection system based on multi-omics deep learning provided in the above embodiments and the colorectal cancer liver metastasis detection method embodiments based on multi-omics deep learning belong to the same concept. For details of its implementation process, please refer to the colorectal cancer liver metastasis detection method embodiments based on multi-omics deep learning, which will not be repeated here.

[0056] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting colorectal liver metastasis based on multi-omics deep learning, characterized in that, Comprising the following steps: Screening a characteristic gene group with CRC metastasis-metabolic reprogramming bifunctionality based on transcriptome dataset, and sample batch effect correction of transcriptomic cohort for deep learning and prediction; Using the characteristic gene group as input, using the characteristic gene group to conduct deep learning on CNN to construct a liver metastasis risk prediction model, calculating the transcriptome risk score based on the metastasis risk probability of the CRC metastasis risk prediction model, and performing correlation analysis of metastatic CRC and CRLM occurrence results according to the transcriptome risk score, screening high correlation features with CRC metastasis and CRLM specificity as initial candidate biomarkers; Obtaining differential expression metabolites corresponding to CRLM patient samples and LCRC patient samples in the metabolic group, performing enrichment pathway analysis on the initial candidate biomarkers corresponding to the transcriptome and the differential expression metabolites corresponding to the metabolic group, constructing a CRLM multi-layer core metabolic network with shared enrichment pathways, and marking the candidate biomarkers in the pathway, using the CRLM multi-layer core metabolic network to test the CRLM resolution performance of the candidate biomarkers, and screening core biomarkers, which are used for precise detection of colorectal cancer liver metastasis.

2. The method for detecting colorectal cancer liver metastases based on multi-omics deep learning according to claim 1, characterized in that, Sample batch effect correction based on transcriptome dataset, including: Merge all transcriptome datasets according to shared genes, use sequencing platform as batch variable and disease status as biological grouping variable, use ComBat for batch effect correction, and retain disease status related biological information, and the correction effect is visualized and verified by PCA, and the average profile coefficient of the sample is calculated ≤ 0.1 to confirm that the batch effect is reduced. 3.The method of claim 1, wherein, Transcriptome risk score calculation based on liver metastasis risk probability of liver metastasis risk prediction model of patient sample characteristic gene group, including: Calculate the average value of the liver metastasis risk probability of the patient sample characteristic gene group in the five sub-models contained in the liver metastasis risk prediction model as the transcriptome risk score. 4.The method of claim 1, wherein, Correlation analysis of metastatic CRC and CRLM occurrence results according to the transcriptome risk score, screening high correlation features with CRC metastasis and CRLM specificity as initial candidate biomarkers, including: Screening high metastasis risk features from the characteristic gene group based on the transcriptome risk score, calculating the correlation coefficient of all features in the mCRC cohort and the CRLM cohort with the transcriptome risk score respectively, and selecting the top N features from the mCRC cohort and the CRLM cohort respectively, and taking the intersection to obtain the core features, which are molecular markers with CRC metastasis and CRLM specificity, and are used as initial candidate biomarkers. 5.The method of claim 1, wherein, The shared enrichment pathway refers to the intersection analysis of the metabolic pathways enriched by the initial candidate biomarkers and the metabolic pathways enriched by the differential expression metabolites, and the key metabolic pathways enriched at the same time, including retinol metabolic pathway and tryptophan metabolic pathway. 6.The method of detecting colorectal cancer liver metastasis based on multi-omics deep learning according to claim 1, characterized in that, The CRLM multi-layer core metabolic network of the common enriched pathway comprises a key metabolic pathway, a biomarker of a regulatory gene, a transcription factor and a metabolite, wherein the transcription factor acts on the biomarker, the biomarker acts on the pathway, and the pathway produces the metabolite; A CRLM core metabolic characteristic map is drawn according to the CRLM multi-layer core metabolic network, and a candidate biomarker is marked in the CRLM core metabolic characteristic map.

7. The multi-omics deep learning-based colorectal cancer liver metastasis detection method of claim 6, wherein, The CRLM multi-layer core metabolic network is used to test the CRLM resolution performance of the candidate biomarker, and a core biomarker is screened, comprising: (1) Performance evaluation of the candidate biomarker: All the candidate biomarkers marked in the CRLM multi-layer core metabolic network are used to distinguish colorectal cancer liver metastasis (CRLM) samples from colorectal cancer primary (CRC PR) samples respectively; the area under the receiver operating characteristic curve of each candidate biomarker is calculated to quantify its distinguishing performance, and the marker with the best distinguishing performance and the highest sensitivity is selected as the candidate marker for CRLM early screening; (2) Performance verification and determination of the core biomarker: The optimal candidate biomarker screened in step (1) is compared with a plurality of reported CRLM prediction biomarkers in terms of performance benchmarking, the distinguishing sensitivity of the candidate marker to CRLM patients is evaluated, the sensitivity difference of the candidate marker to CRLM patients is compared with known CRLM screening markers through ROC curve and AUC value comparative analysis, and the core biomarker is determined according to the evaluation result. 8.The method of claim 1, wherein, Further comprising: The mechanism of the core biomarker in the key metabolic pathway is verified in combination with metabolic monitoring data, and metabolic regulation is realized based on the mechanism, i.e., metabolic reprogramming is performed. 9.A system for detecting colorectal liver metastasis based on multi-omics deep learning, characterized in that, Comprising: A bifunctional feature screening module for screening a feature gene group with CRC metastasis-metabolic reprogramming bifunctionality based on a transcriptome dataset, and performing sample batch effect correction on a transcriptomic cohort for deep learning and prediction; An initial candidate biomarker screening module for using the feature gene group as an input item, using the feature gene group to perform deep learning on a CNN to construct a liver metastasis risk prediction model, calculating a transcriptome risk score based on the metastasis risk probability of the CRC metastasis risk prediction model, and performing correlation analysis of metastatic CRC and CRLM occurrence results according to the transcriptome risk score to screen high-correlation features with CRC metastasis specificity and CRLM specificity as initial candidate biomarkers; A core biomarker screening module for obtaining differentially expressed metabolites corresponding to the metabolic groups of CRLM patient samples and LCRC patient samples, performing enriched pathway analysis on the initial candidate biomarkers corresponding to the transcriptome and the differentially expressed metabolites corresponding to the metabolic groups, constructing a CRLM multi-layer core metabolic network of common enriched pathways, marking the candidate biomarkers in the pathways, using the CRLM multi-layer core metabolic network to test the CRLM resolution performance of the candidate biomarkers, and screening core biomarkers, which are used for precise detection of colorectal cancer liver metastasis.

Images