Multi-modal fusion glioma intelligent prognosis prediction method and system

Abstract

The invention relates to the technical field of image analysis, in particular to a multi-modal fusion glioma intelligent prognosis prediction method and system, and the method comprises the steps: collecting multi-modal data, respectively extracting omics features, constructing a multi-modal fusion depth model, obtaining high and low risk groups, and respectively carrying out difference analysis from multiple omics dimensions, screening out a gene set which is remarkably and highly expressed in the high-risk group, obtaining an intersection of each group of study difference analysis results, preliminarily screening out candidate genes, carrying out inter-study correlation analysis, and screening out intersection genes; introducing the intersection genes into a public database, and screening out genes which are remarkably and highly expressed in GBM to obtain candidate genes; and respectively introducing the candidate genes into a plurality of public databases, and finally identifying potential treatment targets related to the high-risk group. According to the method, the accuracy, the interpretability and the clinical transformation potential of GBM patient layering are comprehensively improved by constructing the multi-modal fusion depth model and combining heterogeneity mechanism analysis and target screening verification.

Description

Technical Field

[0001] This application relates to the field of image analysis technology, and in particular to a multimodal fusion intelligent prognostic prediction method and system for gliomas. Background Technology

[0002] With the rapid development of precision medicine and artificial intelligence technologies, tumor subtyping and prognostic modeling based on multi-omics or multimodal data have gradually become research hotspots. As a highly heterogeneous tumor, glioblastoma involves multiple biological levels in its occurrence and development, including abnormal imaging features, genomic instability, transcriptomic imbalance, and abnormal proteomic expression. Therefore, information from a single modality is insufficient to fully characterize its complexity.

[0003] In recent years, some studies have begun to explore multimodal fusion modeling, attempting to integrate different data sources to improve the accuracy and stability of glioblastoma (GBM) risk assessment. Existing research mainly focuses on joint modeling of radiomics and clinical data, such as constructing simple survival prediction models based on MRI imaging features and clinical indicators such as age and KPS scores. Another common approach is based on transcriptome data, constructing gene feature risk scoring models for prognostic stratification. However, these methods often employ simple data splicing or weighting strategies, ignoring the potentially complex interactions between different modalities and failing to fully utilize complementary information from multiple omics.

[0004] However, these methods typically achieve only shallow integration, lack a unified and standardized data processing workflow, have insufficient fusion depth, and lack systematic follow-up biological validation and potential target discovery. Therefore, although multimodal fusion is becoming a research trend, there is currently a lack of a mature technical solution that can simultaneously achieve systematic breakthroughs in data fusion depth, model survival prediction performance, biological mechanism interpretability, and therapeutic target discovery. This is precisely the problem that this application aims to solve. Summary of the Invention

[0005] This application provides a multimodal fusion-based intelligent prognostic prediction method and system for gliomas. By constructing a multimodal fusion deep model and combining heterogeneity mechanism analysis and target screening verification, it comprehensively improves the accuracy, interpretability, and clinical translation potential of GBM patient stratification, and systematically solves the core technical problems of low accuracy, poor interpretability, and lack of treatment intervention guidance in existing single-omics modeling.

[0006] To address the aforementioned technical problems, in a first aspect, embodiments of this application provide a multimodal fusion-based intelligent prognostic prediction method for gliomas, comprising the following steps: First, collecting multimodal data and extracting omics features respectively; then, constructing a multimodal fusion deep model based on deep learning algorithms and omics features; next, based on the multimodal fusion deep model, obtaining high- and low-risk groups, and performing differential analysis from the dimensions of genome, copy number variation, transcriptome, proteome, and immune microenvironment to obtain the results of differential analysis of each omics; then, based on the results of differential analysis of each omics, screening out a set of genes that are significantly highly expressed in the high-risk group, and obtaining the intersection of the results of differential analysis of each omics, initially screening out candidate genes that are significantly related to the high-risk group, and performing inter-omics correlation analysis to screen out intersection genes that are significantly correlated and have consistent trends; next, introducing the intersection genes into a public database, screening out genes that are significantly highly expressed in GBM to obtain candidate genes; finally, introducing the candidate genes into multiple public databases respectively, and combining survival analysis and statistical indicators to ultimately identify potential therapeutic targets related to the high-risk group.

[0007] In some exemplary embodiments, the process of acquiring multimodal data and extracting omics features includes: acquiring multimodal data, standardizing the multimodal data, and extracting omics features corresponding to each omics from the standardized multimodal data; wherein the multimodal data includes magnetic resonance imaging, whole tissue section scanning, exome sequencing, transcriptome sequencing, and proteomics data; and the omics features include radiomics features, pathomics features, whole exome sequencing data, transcriptome sequencing data, and proteomics features.

[0008] In some exemplary embodiments, a multimodal fusion deep model is constructed based on deep learning algorithms and the omics features, including: performing multimodal data fusion on the omics features corresponding to each omics, and constructing a multimodal fusion hierarchical model based on deep learning algorithms.

[0009] In some exemplary embodiments, differential analyses are performed from the dimensions of genome, copy number variation, transcriptome, proteome and immune microenvironment to reveal biological differences among patients in different risk groups; at the genome level, the differences in gene mutation types, gene variant types and single base substitution types among patients in different risk groups are compared based on the Wilcoxon rank-sum test.

[0010] In some exemplary embodiments, at the copy number variation level, the limma method is used to identify chromosomal copy number variation segments with significant differences in different risk groups; at the transcriptomic and proteomic levels, the Wilcoxon rank-sum test is used to screen genes and proteins with significant differences in expression levels in high- and low-risk group samples, with the screening criteria being FDR < 0.05 and log2FC absolute value greater than 1, to screen differentially expressed genes and proteins; then, ORA analysis is performed on the screened differentially expressed genes and proteins, and the enriched pathways are functionally clustered and visualized.

[0011] In some exemplary embodiments, at the level of immune microenvironment feature analysis, immune-related pathways involved in gene set enrichment analysis (GSEA) are screened out, and their activation status differences are classified and visualized. Then, based on the ESTIMATE algorithm, the infiltration degree of stromal cells and immune cells in tumor tissue is estimated, and the differences between immune scores and stromal scores between the two risk groups are analyzed.

[0012] In some exemplary embodiments, when performing inter-omics correlation analysis, the correlation between copy number variation and transcriptome expression, as well as the correlation between transcriptome expression and proteome expression, is calculated to screen out intersection genes with significant correlations and consistent trends as key genes, so as to ensure the biological consistency of the screened genes at different omics levels.

[0013] In some exemplary embodiments, the intersection genes are introduced into a public database to screen for genes that are significantly overexpressed in GBM and obtain candidate genes. This includes: introducing the intersection genes into the TCGA and CGGA databases, and screening for genes that are significantly overexpressed in GBM based on the difference in transcriptomic expression between GBM and low-grade glioma, so as to enhance the tumor specificity of the candidate genes.

[0014] In some exemplary embodiments, candidate genes are introduced into multiple public databases, and survival analysis and statistical indicators are combined to finally identify potential therapeutic targets related to the high-risk group. This includes: introducing candidate genes into multiple public databases, combining survival analysis, calculating the P-value of the Log-rank test and the P-value of the univariate Cox regression, screening out genes whose survival analysis P-values ​​at multiple omics levels are all less than 0.05, and counting and ranking them according to statistical indicators to finally identify potential therapeutic targets related to the high-risk group.

[0015] Secondly, this application also provides a multimodal fusion intelligent prognostic prediction system for gliomas. This system is used to implement the multimodal fusion intelligent prognostic prediction method for gliomas as described in the above embodiments. The system includes: a multimodal data acquisition module, a model building module, a heterogeneity analysis module, and a therapeutic target screening module connected in sequence. The multimodal data acquisition module collects multimodal data and extracts omics features respectively. The model building module constructs a multimodal fusion deep model based on a deep learning algorithm and the omics features. The heterogeneity analysis module obtains high- and low-risk groups based on the multimodal fusion deep model, respectively from the perspectives of genome, copy number variation, transcriptome, and protein. Differential analysis was performed on the white matter and immune microenvironment dimensions to obtain the results of differential analysis of each omics. The therapeutic target screening module was used to screen the set of genes that were significantly highly expressed in the high-risk group based on the results of differential analysis of each omics, and to obtain the intersection of the results of differential analysis of each omics. Candidate genes that were significantly related to the high-risk group were initially screened, and inter-omics correlation analysis was performed to screen the intersection genes that were significantly correlated and had consistent trends. The intersection genes were introduced into a public database to screen the genes that were significantly highly expressed in GBM to obtain candidate genes. The candidate genes were introduced into multiple public databases respectively, and combined with survival analysis and statistical indicators, the potential therapeutic targets related to the high-risk group were finally identified.

[0016] The technical solution provided in this application has at least the following advantages: This application provides a multimodal fusion-based intelligent prognostic prediction method and system for gliomas. The method includes the following steps: First, multimodal data is collected, and omics features are extracted respectively. Then, a multimodal fusion deep model is constructed based on deep learning algorithms and omics features. Next, based on the multimodal fusion deep model, high- and low-risk groups are obtained, and differential analysis is performed from the dimensions of genome, copy number variation, transcriptome, proteome, and immune microenvironment to obtain the differential analysis results of each omics. Then, based on the differential analysis results of each omics, a set of genes that are significantly highly expressed in the high-risk group is screened, and the intersection of the differential analysis results of each omics is obtained. Candidate genes that are significantly related to the high-risk group are initially screened, and inter-omics correlation analysis is performed to screen the intersection genes that are significantly correlated and have consistent trends. Next, the intersection genes are introduced into a public database to screen genes that are significantly highly expressed in GBM to obtain candidate genes. Finally, the candidate genes are introduced into multiple public databases, and combined with survival analysis and statistical indicators, potential therapeutic targets related to the high-risk group are finally identified.

[0017] This application aims to overcome the shortcomings of existing multimodal modeling techniques in terms of fusion depth, data standardization, interpretability validation, and clinical translation applications. It proposes a systematic, complete, and scalable new method for GBM risk stratification and target discovery based on multimodal data fusion, achieving an organic unity between accurate risk prediction and personalized treatment intervention. By constructing a multimodal fusion depth model, combined with heterogeneity mechanism analysis and target screening validation, this application comprehensively improves the accuracy, interpretability, and clinical translation potential of GBM patient stratification. Attached Figure Description

[0018] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations do not constitute a limitation on the embodiments, and unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0019] Figure 1 This is a flowchart illustrating a multimodal fusion-based intelligent prognostic prediction method for gliomas, provided as an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of the architecture of a multimodal fusion intelligent prognostic prediction system for glioma, provided in one embodiment of this application. Detailed Implementation

[0021] As the background technology shows, existing methods usually only achieve shallow integration, lack a unified standardized data processing process, have insufficient fusion depth, and lack systematic follow-up biological validation and potential target discovery.

[0022] Currently, GBM is the most common and most aggressive central nervous system tumor, and its clinical treatment and prognostic management face enormous challenges. Traditional risk stratification methods often rely on a single modality (such as MRI images, gene mutation status, or clinical parameters). Because they cannot fully characterize the complex biological characteristics of GBM, the accuracy and stability of stratification are insufficient, making it difficult to support precise treatment decisions.

[0023] The closest existing approach to this application is a partial "bimodal" fusion study, such as combining radiomics and genomics for mutation prediction, or fusing clinical data with transcriptome features for survival prediction. However, these methods typically achieve only a shallow level of integration, lack a unified standardized data processing workflow, have insufficient fusion depth, and lack systematic follow-up biological validation and potential target discovery. Furthermore, the methodologies for validating the biological consistency of multi-omics studies (such as validating the rationality of fusion grouping across multiple dimensions like gene mutation, pathway enrichment, and immune infiltration levels) are still immature. The utilization of emerging omics technologies such as single-cell omics and spatial omics is also very limited, making it difficult for existing multimodal models to balance survival prediction performance with the interpretability of biological mechanisms.

[0024] The shortcomings of existing technologies are mainly reflected in the following aspects: First, existing technologies generally suffer from a lack of simplistic fusion strategies and limited model performance. Most multimodal fusion methods employ simple feature concatenation or weighted averaging, failing to fully consider the interactions and complementarities between features from different omics systems. This results in survival prediction models with limited accuracy, poor robustness, and susceptibility to specific modal noise.

[0025] Secondly, existing methods lack standardized processes for processing and integrating multi-omics data. The quality of data from different omics varies, and the fusion effect is uncontrollable, which further limits the reproducibility and application of the model.

[0026] Third, existing multimodal fusion models generally lack sufficient biological interpretability and systematic analysis of heterogeneous mechanisms. Most studies only statistically validate the model's survival prediction ability without exploring the differences in molecular mechanisms behind different risk groups.

[0027] Fourth, existing studies generally lack systematic mining and functional validation of potential therapeutic targets in risk groups, resulting in models that, while capable of stratification, cannot guide subsequent precision treatment decisions.

[0028] Traditional risk stratification methods rely on single-modality techniques. The primary technical problem this application aims to solve is how to achieve higher accuracy and stability in risk stratification of GBM patients based on multimodal data fusion.

[0029] Secondly, existing methods based on single-omics or simple bimodal fusion often overlook the potential interactions and complementary information between different omics data, leading to overfitting or poor generalization in risk assessment models. The second technical problem this application aims to solve is: how to fully explore the synergistic effects between radiomics, pathomics, genomics, transcriptomics, and proteomics data through a scientifically sound deep fusion strategy, thereby constructing a more robust and generalizable survival prediction model.

[0030] Furthermore, existing multimodal models generally lack systematic biological mechanism validation, making it difficult to explain the biological rationale behind model stratification, thus affecting clinicians' trust in model results and their willingness to apply them. Therefore, the third technical problem that this application aims to address is: how to establish a systematic multi-omics biological validation framework to ensure that stratification results have clear biological interpretability and consistency at the molecular, cellular, and tissue levels.

[0031] Existing research rarely focuses on directly linking risk stratification results with potential therapeutic targets, resulting in stratification models possessing prognostic predictive capabilities but lacking the ability to guide therapeutic interventions. This application addresses this gap by proposing a solution to the fourth technical problem: how to discover risk-group-related therapeutic targets through fusion models and functionally validate them in in vitro and in vivo models, thereby achieving a complete closed loop from risk prediction to therapeutic intervention. In summary, this application aims to comprehensively improve the accuracy, interpretability, and clinical translational potential of GBM patient stratification by constructing a multimodal fusion deep model, combined with heterogeneity mechanism analysis and target screening validation, systematically solving the core technical problems of low accuracy, poor interpretability, and lack of therapeutic intervention guidance in existing single-omics modeling.

[0032] The embodiments of this application will now be described in detail with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the embodiments of this application to facilitate a better understanding of the application. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0033] See Figure 1 This application provides a multimodal fusion-based intelligent prognostic prediction method for gliomas, comprising the following steps: Step S1: Collect multimodal data and extract omics features respectively.

[0034] Step S2: Construct a multimodal fusion deep model based on deep learning algorithms and omics features.

[0035] Step S3: Based on the multimodal fusion deep model, high- and low-risk groups are obtained, and differential analysis is performed from the dimensions of genome, copy number variation, transcriptome, proteome and immune microenvironment to obtain the differential analysis results of each omics.

[0036] Step S4: Based on the results of differential analysis of each omics, screen out the set of genes that are significantly overexpressed in the high-risk group, and obtain the intersection of the results of differential analysis of each omics. Preliminarily screen out candidate genes that are significantly related to the high-risk group, and perform inter-omics correlation analysis to screen out intersection genes that are significantly correlated and have consistent trends.

[0037] Step S5: Introduce the intersecting genes into a public database, screen out genes that are significantly overexpressed in GBM, and obtain candidate genes.

[0038] Step S6: The candidate genes are introduced into multiple public databases, and combined with survival analysis and statistical indicators, potential therapeutic targets related to the high-risk group are finally identified.

[0039] This application addresses the problems of low accuracy and poor stability in existing glioma prognostic prediction methods, which rely on single-modal data, and the shallow nature of existing multimodal fusion strategies, which neglect data interactions and have insufficient model generalization ability. Furthermore, existing methods lack systematic biological validation and have weak interpretability; they also struggle to correlate with effective therapeutic targets and thus cannot guide clinical intervention. Therefore, this application provides a multimodal fusion-based intelligent prognostic prediction method for gliomas. It employs a five-modal deep fusion strategy, integrating multimodal data such as MRI images, pathology, genomics, transcriptomics, and proteomics to construct a high-precision, robust deep learning model for risk stratification. Combined with multi-omics heterogeneity analysis and therapeutic target screening validation, this significantly improves the accuracy, robustness, and clinical translational value of prognostic prediction, achieving a closed loop from risk prediction to precision treatment intervention.

[0040] Specifically, in some embodiments, step S1 involves collecting multimodal data and extracting omics features, including: acquiring multimodal data, standardizing the multimodal data, and extracting the omics features corresponding to each omics from the standardized multimodal data.

[0041] Multimodal data include magnetic resonance imaging, whole tissue section scanning, exome sequencing, transcriptome sequencing, and proteomics data; the omics features include radiomics features, pathomics features, whole exome sequencing data, transcriptome sequencing data, and proteomics features.

[0042] Specifically, step S1, which involves extracting radiomics features from multimodal data, includes the following steps: First, MRI images are extracted from the multimodal data; MRI images are magnetic resonance imaging images.

[0043] Then, the MRI images are standardized, and imaging features are extracted from the standardized MRI images to obtain radiomics features.

[0044] Specifically, during the MRI scanning and imaging feature extraction process, the patient's MRI images were obtained using 3.0 T MRI scans during routine examinations. Sequences included: axial and sagittal T1-weighted imaging (T1WI), axial T2-weighted imaging (T2WI), axial T2-weighted fluid attenuation inversion recovery (FLAIR) imaging, and axial, sagittal, and coronal contrast-enhanced T1-weighted imaging (CE-T1WI) immediately after intravenous injection of a 0.1 mmol / kg gadolinium contrast agent. Apparent diffusion coefficient (ADC) maps were obtained using axial diffusion-weighted imaging (DWI). Acquisition parameters for each sequence were as follows: (1) T1WI and CE-T1WI: Repetition time (TR) 220~1750 ms; Echo time (TE) 2.3~24 ms; Echo sequence length (ETL) 1~12; Slice thickness 5 mm; Average value / excitation 1; Flip angle (FA) 70~111; Field of view (FOV) 220×192~240×240 mm; Matrix 256×162-320×256 mm.

[0045] (2) T2WI: TR 1873~5390 ms; TE 70~117 ms; ETL 16~32; slice thickness 5 mm; average value / excitation 1; FA 90~142; FOV 220×192-240×240 mm; matrix 320×238-512×512 mm.

[0046] (3) FLAIR: TR 4500~8400 ms; TE 85~150 ms; Reversal time (TI) 1670~2250 ms; ETL1-38; Slice thickness 5 mm; Average value / excitation 1; FA 90~150; FOV 220×192~240×240 mm; Matrix 256×179~256×256 mm.

[0047] (4) DWI: The images are processed by the corresponding post-processing workstation. The ADC images are calculated from DWI acquired at B-values ​​of 0 and 1000 s / mm. The sequence parameters include: TR 2121~6000 ms; TE 77~119 ms; ETL 1~82; slice thickness 5 mm; mean / excitation 1; FA 90 FOV 220×220~240×240 mm; matrix 152×114~192×192 mm. The ADC map of all imaging planes is generated on a voxel-by-voxel basis using a single exponential model.

[0048] First, the N4ITK algorithm was used to correct bias field distortion. After resampling isotropic voxels to 1×1×1 mm using trilinear interpolation, axially resampled CE-T1WI images were used as templates, and mutual information was used as a similarity metric for rigid registration of multiple MRI sequences for each patient. This process was performed using 3D Slicer software, generating registered images rT1WI, rCE-T1WI, rT2WI, rFLAIR, and rADC. Gray-level normalization was performed using histogram matching. A neuroimaging associate chief physician with over 10 years of experience in head MRI diagnosis manually delineated the region of interest (ROI) on the axial planes of rFLAIR, rT2WI, and rCE-T1WI images using ITK-SNAP software, obtaining the volume of interest (VOI). VOI was defined as the enhancing, non-enhancing, and necrotic areas of the tumor. VOI contours were drawn based on FLAIR images, and the tumor extent was cross-checked using rT2WI and rCE-T1WI images, with fine-tuning of the tumor contour. The radiologist and a neurosurgeon with over 10 years of experience used a simple random sampling method to randomly select 100 patients from the group for VOI re-images. The inter-group correlation coefficient (ICC) was used to evaluate the intra-rater reliability analysis of the test-retest dataset and the inter-rater reliability analysis of the multi-description dataset, preserving features with ICC ≥ 0.75. The obtained VOI images were then overlaid with co-registered rT1WI, rCE-T1WI, rT2WI, rFLAIR, and rADC images.

[0049] Three types of features were extracted using a pyramid model: first-order intensity statistics, shape descriptors, and higher-order texture features. Five basic matrices were used to define texture features: gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size region matrix (GLSZM), gray-level dependence matrix (GLDM), and neighborhood gray-level difference matrix (NGTDM). Imaging features were extracted from three types of images: original images, wavelet images, and Laplacian Gaussian images. Ultimately, 5929 features were extracted from five MRI sequences, with 4271 features retaining an ICC ≥ 0.75.

[0050] In some embodiments, step S1, extracting pathomic features from multimodal data, includes the following steps: First, WSI images are obtained by scanning a hematoxylin-eosin slide; WSI images are full slide images.

[0051] Then, tissue segmentation was performed on the WSI images, and features were extracted from the segmented tissue images to obtain pathomic features.

[0052] Specifically, during the acquisition and feature analysis of WSI images, a MAGSCAN-NER scanner (KF-PRO-005, KFBIO) was used to scan hematoxylin and eosin (H&E) slides to obtain WSI images (WSIs). In WSIs, tissue typically occupies a portion of the white background on the slide, thus requiring initial tissue segmentation of the WSI images. The WSIs at 5x resolution were converted from RGB to Lab color space, and the Otsu algorithm was used to calculate the tissue segmentation threshold. The segmented tissue images were then divided into numerous 1024×1024 patches at 20x objective magnification (0.5m / pixel), and feature extraction was performed using CellProfiler. A total of 1035 features were extracted from 122 patients.

[0053] In some embodiments, step S1, which involves extracting whole-exome sequencing (WES) data and transcriptome sequencing data from multimodal data, includes performing whole-exome sequencing and transcriptome sequencing on tumor samples in the multimodal data to obtain whole-exome sequencing data and transcriptome sequencing data, which are denoted as WES data and RNA-seq data, respectively.

[0054] Specifically, during WES and analysis, DNA from tumor tissue and adjacent brain tissue was extracted from samples using the QIAamp Rapid DNA Tissue Kit (Qiagen). Blood samples were collected in tubes containing EDTA and centrifuged at 1600 xg for 10 minutes at 4°C within 2 hours of collection. Peripheral blood lymphocyte (PBL) particles were stored at -20°C until further use, and PBL DNA was extracted using the Relax Gene Blood DNA System. DNA quantification was performed using a Quantum Bit 3.0 fluorometer and the Quantum Bit dsDNA HS Analysis Kit. DNA from tissue and PBL samples was fragmented using dsDNA fragmentation enzyme, followed by size selection of DNA fragments (150–250 bp) using Ampure XP magnetic beads. DNA fragment libraries were constructed using the KAPA Library Preparation Kit. Cleaning steps were performed using Agencourt AMPure XP magnetic beads. After DNA fragmentation, end repair and 3'A tailing were performed, followed by exon capture using the Agilent SureSelect Human Whole Exon V6 Kit. The purity and concentration of DNA fragments were assessed using a QuantumBit 3.0 fluorometer and a QuantumBit dsDNA HS analysis kit. Fragment lengths were measured on a 4200 bioanalyzer using the DNA1000 kit. DNA libraries with 150 bp terminal sequences were sequenced using an Illumina Novaseq 6000 system. Raw data were converted to FASTQ files, and adapters and low-quality reads were trimmed using trim atorial. The median coverage depth was 112x for tumor specimens and 128x for non-tumor specimens.

[0055] Single nucleotide variants (SNVs) and insertions or deletions (INDELs) were identified using GATK tools. Paired-end WES reads were mapped to the human reference genome (hg38) using BWA-mem. BAM files were further processed using Picard through reordering, sorting, tagging duplicates, and adding read groups. Basic quality score recalibration was performed using the Base Recalibrator module in GATK, followed by evaluation of cross-sample contamination using the Get Pileup Summaries and Calculate Contamination modules. Somatic variant 2 was detected by mutation detection, using ANNOVAR annotation with patient-matched normal DNA sequencing reads as a reference. Candidate somatic variants were differentiated according to the following screening criteria: ① variants other than exon regions and splice sites were excluded; ② variants with a variant allele fraction (VAF) ≥ 5% in the tumor sample and at least two supporting variant reads were retained; ③ at least one variant with a mutational allele frequency (MAF) ≥ 5% in the database, including 1000 genomes, ESP6500, gnomAD, and ExAC, was removed. Normal samples were sequenced using the same protocol, reducing each sample to 4% and then merging them as a reference. To obtain high-quality and reliable somatic variants, this application employed stringent downstream filtering criteria: ① variants other than exon regions and splicing sites were excluded; ② variants with VAF ≥ 5%, at least 5 supporting variant reads in tumor samples, and tumor VAF variants with VAF ≥ five times that of normal samples were retained; ③ variants appearing more than 100 times in COSMIC (v92) were retained; ④ variants with MAF ≥ 1% in at least one variant database (1000 genomes, ESP6500, gnomAD, and ExAC) were removed; ⑤ variants predicted as benign by at least two of the following tools: mutation vectors, mutation containers 2, polyphenols 2, and SIFT were removed. The CNVkit inferred somatic CNVs based on the BAM file generated during somatic mutation detection using the default cyclic binary segmentation algorithm. The log2 ratio at the fragment level was calculated and converted into input for the GISTIC 2.0 software to identify significantly amplified or deleted chromosomal regions in the tumor. CNV amplification and deletion are defined using a 0.3 log2 ratio threshold.

[0056] Total RNA was extracted from tissue samples using the TRIzol kit during transcriptome sequencing (RNA-seq) and analysis. RNA concentration and integrity were assessed using the Qubit RNA Analysis Kit, Qubit 2.0 fluorometer, and Agilent 2100 Bioanalyzer. Samples with RNA integrity values ​​greater than 5 were included in the study. Libraries were prepared from samples with high RNA integrity, no contaminants, and sufficient RNA quantity. RNA was purified from total RNA using poly-T oligonucleotide magnetic beads. RNA was lysed at high temperature using divalent cations in NEBNext first-strand synthesis reaction buffer (5X). cDNA synthesis, end repair, A-tailing, and NEBNext adaptor ligation were performed using the NEBNext Ultra RNA Library Preparation Kit. Library fragments were purified using AMPure XP, selecting cDNA fragments of 150–200 bp in length. Library quality was assessed using an Agilent Bioanalyzer. The libraries were sequenced on the Illumina HiSeq X Ten platform, producing 150 bp paired-end reads. Sequencing data were filtered using the trimcardio software to remove adaptors and low-quality sequences, and then FastQC was used for data quality assessment. Sequences were aligned to a reference genome (hg38) using STAR. Gene expression values ​​were calculated using RSEM based on the GENCODE (v35) gene annotation file. HTSeq was used to count the number of reads aligned to each gene, and gene expression levels were quantified as FPKM (fragments per thousand bases per million mapped exons) and TPM (transcriptions per thousand bases per million mapped exons).

[0057] In some embodiments, step S1, which involves extracting proteomics features from multimodal data, includes performing mass spectrometry analysis on tumor samples in the multimodal data to obtain proteomics features.

[0058] Specifically, mass spectrometry analysis was performed on tumor samples from the multimodal data. The procedure included: removing samples from a -80°C storage chamber, weighing an appropriate amount of paper towel, and placing it in a liquid nitrogen-pre-cooled mortar. Liquid nitrogen was added, and the tissue was thoroughly ground into powder. Four times the powder volume of lysis buffer (1% Triton X-100, 1% protease inhibitor, 1% phosphatase inhibitor, 3 μm TSA, 50 mM NAM) was added to each sample, followed by sonication lysis. The samples were centrifuged at 12,000 g for 10 minutes at 4°C to remove cell debris, and the supernatant was transferred to a new centrifuge tube. Protein concentration was determined using a BCA assay kit.

[0059] Digest an equal volume of protein in each sample with trypsin and adjust the volume using lysis buffer. Add one part pre-cooled acetone and vortex, then add four parts pre-cooled acetone and precipitate at -20°C for 2 hours. Centrifuge the sample at 4,500 g for 5 minutes and discard the supernatant. Wash the precipitate twice with pre-cooled acetone. After air-drying the precipitate, resuspend it in a 200 mM TEAB vial and add trypsin at a 1:50 ratio (protease:protein, w / w) for overnight digestion. Add dithiothreitol (DTT) to a final concentration of 5 mM and reduce the sample at 56°C for 30 minutes. Add iodoacetamide (IAA) to a final concentration of 11 mM and incubate the sample in the dark at room temperature for 15 minutes.

[0060] Samples were separated using an Agilent 300 Extend C18 column (4.6 × 250 mm) at a detection wavelength of 214 nm and a column temperature of 35°C. The column was equilibrated with 95% buffer A for 30 minutes. After baseline stabilization, a fractional gradient method was initiated, loading the peptide samples into a high-performance liquid chromatography (HPLC) column. Samples were collected every 1 minute, and fractions 11 to 46 were combined into 12 groups and vacuum dried. The peptides were dissolved in mobile phase A and separated using an EASY-nLC 1200 ultra-high performance liquid chromatography system. Mobile phase A consisted of 0.1% formic acid and 2% acetonitrile aqueous solution, and mobile phase B consisted of 0.1% formic acid and 90% acetonitrile aqueous solution. The gradient was set as follows: 0–96 min, 6%–25% B; 96–114 min, 25%–35% B; 114–117 min, 35%–80% B; and 117–1200 min, 80% B, with the flow rate maintained at 500 nL / min. The isolated peptides were ionized in an NSI ion source, and data were collected using an Orbitrap Exploris 480 mass spectrometer.

[0061] The liquid chromatography (LC) parameters were consistent with those used during library construction. Peptides were separated using an ultra-high performance liquid chromatography (UHPLC) system and analyzed using an Orbitrap Exploris 480 mass spectrometer. Precursor ions and their fragment ions were detected and analyzed using a high-resolution Orbitrap. The FAIMS compensation voltage (CV) was set to -40 V, -55 V, and -70 V. The primary mass scan range was set to 350–1350 m / z with a resolution of 120,000; the secondary scan resolution was set to 30,000. Secondary data acquisition mode was set to DIA mode, followed by a primary scan where peptide ions in a 20 m / z window were fragmented in an HCD collision cell using 32% collision energy, followed by secondary mass analysis. Automatic gain control (AGC) for the secondary spectrum was set to 600%.

[0062] It should be noted that this application aims to collect as much glioblastoma multiforme (GBM) sequencing data as possible from public databases to validate conclusions and enrich the research content. This application obtained transcriptome expression profiles and clinical data of samples from the Gene Expression Comprehensive Database and the Chinese Glioma Genome Atlas, including GSE72951 (GPL14951, n=112), GSE43289 (GPL570, n=26), GSE43378 (GPL570, n=32), GSE7696 (GPL570, n=84), GSE13041 (GPL570 and GPL96, n=218), GSE15824 (GPL570, n=25), and GS from a series of matrix files uploaded by the authors. This application used the `normalizeBetweenArrays` function in the limma software package to perform quantile normalization on the microarray data. Subsequently, after eliminating batch effects, this application merged datasets from the same sequencing platform, resulting in cohorts of GBM-GPL570 (n=215), GBM-GPL6480 (n=110), GBM-GPL96 (n=326), and GBM-GPL97 (n=135). The combat function in the sva package was used to remove batch effects. Finally, each dataset was normalized using a scaling function.

[0063] This application employs unsupervised clustering-based multimodal fusion. Integrating multimodal data can reveal causal features that may be masked in unimodal analysis and provide a holistic understanding of disease complexity by exploring the interactions between modalities and how these relationships drive differences in patient outcomes (such as survival and drug response). Multimodal data fusion strategies can be categorized into early fusion, mid-fusion, and late fusion based on their time frame. Early fusion connects all data forms into a single matrix, which can lead to the "curse of dimensionality" and variable shifts in subsequent analyses and fails to correct imbalances in multimodal data, potentially adversely affecting downstream analyses. Late fusion involves analyzing each omics layer separately and then integrating the results to produce consistent outcomes and outputs. However, this approach sacrifices complementary interaction information from the multimodal data. Mid-fusion typically involves simultaneous integration and clustering to connect correlations between different omics layers, identify multimodal joint clusters, and infer patient stratification and molecular mechanisms. Generally, mid-fusion is more advanced but requires sophisticated fusion algorithms.

[0064] In some embodiments, multimodal data fusion of radiomics features, pathomics features, whole-exome sequencing data, transcriptome sequencing data, and proteomics features includes the following steps: First, multiple algorithms based on different principles are used to perform intermediate fusion of multimodal data, including radiomics features, pathomics features, whole exome sequencing data, transcriptome sequencing data, and proteomics features.

[0065] Then, the intermediate fusion results obtained from multiple algorithms are fused in the later stage to obtain the final clustering result.

[0066] This application integrates multiple algorithms based on different principles to perform intermediate fusion of multimodal data (FAHZZU1 queue), and then performs late-stage fusion of the results obtained from the multiple algorithms to obtain the final clustering result. Feature matrices from whole-exome sequencing (WES), transcriptomics (RNA-seq), proteomics (LC-MS), pathomics (WSI), and radiomics data are fused using multiple different algorithms in the intermediate stage, and then consensus clustering is generated through late-stage fusion.

[0067] In some embodiments, prior to intermediate fusion of multimodal data including radiomics features, pathomics features, whole-exome sequencing data, transcriptome sequencing data, and proteomics features, the method further includes: determining the optimal number of clusters and input features.

[0068] Specifically, the process for determining the optimal number of clusters and input features includes: (1) First, the median absolute deviation (MAD) of variables in each modality layer of radiomics, pathomics, transcriptomics, and proteomics is calculated. The top n variables from each layer are selected and combined into 2640 variable combinations. Then, the cluster prediction index (CPI) and GAP statistic of each combination are calculated. Based on the combination with the highest sum of CPI and GAP statistics, the optimal number of clusters and input features for the final clustering are determined. The CPI and GAP statistics are calculated to determine the optimal number of clusters, with K=3 as the optimal number of clusters. (2) Eleven algorithms based on different principles were used for intermediate fusion of multimodal data. These algorithms include CIMLR, CPCA, iClusterBayes, IntNMF, LRAcluster, MCIA, NEMO, PINSPlus, RGCCA, SGCCA and SNF.

[0069] (3) The Jaccard index was calculated using the binary results of 11 algorithms to assess the similarity between samples.

[0070] (4) Clustering analysis is used to obtain consistent results from 11 algorithms based on the Jaccard distance matrix.

[0071] (5) The proportion of fuzzy clustering (PAC) and the Calinski and Harabasz index (CHI) are used to evaluate the fitness of the number of clusters.

[0072] (6) The silhouette coefficient is calculated for each cluster, and samples with a silhouette coefficient lower than 0.4 are removed to obtain the core sample set.

[0073] In some embodiments, step S2, which constructs a multimodal fusion deep model based on deep learning algorithms and omics features, includes: performing multimodal data fusion on the omics features corresponding to each omics, and constructing a multimodal fusion hierarchical model based on deep learning algorithms.

[0074] After the model was built, this application conducted a systematic multi-omics heterogeneity analysis based on the high and low risk groups identified by the model, aiming to reveal the biological differences of patients in different risk groups by systematically integrating genomic, copy number variation, transcriptomics, proteomics and immune microenvironment data.

[0075] In some embodiments, step S3 involves differential analysis at the genomic, copy number variation, transcriptomic, proteomic, and immune microenvironment dimensions to reveal biological differences among patients in different risk groups. At the genomic level, based on the Wilcoxon rank-sum test, differences in gene mutation types (including frameshift mutations, non-frameshift mutations, point mutations, splicing mutations, and translation initiation mutations), gene variation types (including deletion mutations, insertion mutations, polynucleotide polymorphisms, and single nucleotide polymorphisms), and single base substitution types (including C>A, C>T, T>A, C>G, T>C, and T>G) are compared among patients in different risk groups. Further, the mutation frequency differences of each gene among different groups are statistically analyzed, characteristic genes with a mutation frequency greater than 5% are screened, and pathway enrichment analysis (ORA) is performed using the ClusterProfiler software package to screen for biological pathways with a p-value less than 0.05. In addition, two groups of specific tumor-driving genes are screened using the OncodriveCLUST algorithm, and ORA analysis is used to further reveal related pathways. This application also identifies gene pairs with shared or mutually exclusive mutation relationships among risk groups using Fisher's exact test, visualizes their characteristic gene networks using the GGally package, and performs functional annotation and enrichment analysis on the relevant genes. Finally, based on the Reactome, KEGG, and Hallmark databases, this application calculates the mutational burden differences of each pathway and screens out key pathways affecting more than 10% of patients, further revealing the potential impact of inter-group mutational characteristics on biological function.

[0076] In some embodiments, at the copy number variation (CNV) level, this application uses the limma method to identify significantly different chromosomal copy number variation segments in different risk groups based on the results of GISTIC 2.0 analysis. Further, the gene sets corresponding to these significant segments are extracted, and ORA functional enrichment analysis is performed using ClusterProfiler based on copy amplification or deletion, screening for key pathways with p-values ​​less than 0.05, and analyzing the impact of CNV events on biological pathway function.

[0077] In some embodiments, at the transcriptomic and proteomic levels, this application uses the Wilcoxon rank-sum test to screen genes and proteins with significantly different expression levels in high- and low-risk group samples. The screening criteria are FDR < 0.05 and absolute log2FC value greater than 1. Further, ORA analysis is performed on the screened differentially expressed genes and proteins, and the enriched pathways are functionally clustered and visualized using the aPEAR software package. Simultaneously, based on the log2FC ranking of all differentially expressed genes, this application uses the GSEA method to screen enriched pathways with FDR < 0.05. Combined with GSVA analysis results, the activity changes of relevant pathways in each patient sample in the Reactome, KEGG, and Hallmark databases are scored and displayed in heatmaps, thereby systematically revealing the differences in expression characteristics at the functional pathway level among different risk groups.

[0078] To analyze the characteristics of the immune microenvironment, this application identified immune-related pathways involved in the GSEA analysis and classified and visualized the differences in their activation states. Furthermore, the infiltration levels of stromal cells and immune cells in tumor tissue were estimated using the ESTIMATE algorithm, and the differences in immune scores and stromal scores between the two risk groups were analyzed. Simultaneously, differences in four types of immune function activities—antigen presentation (MHC), effector cells (EC), suppressor cells (SC), and immune checkpoints (CP)—were assessed. This application also used transcriptomics to infer the sensitivity of tumors to immune checkpoint inhibitors (such as PD-1 / PD-L1) therapy and compared the differences in anti-cancer immune function scores across seven stages of the immune cycle, systematically characterizing the heterogeneity of immune function status between the two patient groups.

[0079] To further validate the heterogeneity of the risk grouping model at the cellular level, this application constructs a classifier based on a fusion multimodal cohort. Logistic regression combined with Lasso regression is used for feature screening, retaining important genes with non-zero coefficients as input variables. A classification model is established using 14 machine learning algorithms (AdaBoost, DecisionTree, ElasticNet, GBDT, KNN, Lasso, LDA, NBayes, NNet, RF, Ridge, StepLR, SVM, and XGBoost) to predict patient risk grouping in single-cell pseudo-bulk data. Furthermore, based on the grouping results, the Wilcoxon rank-sum test is used to analyze differences in cell subpopulation proportions and screen for significantly altered cell subtypes. Differentially expressed genes are screened within each cell subpopulation and functionally enriched to analyze changes in functional characteristics. Simultaneously, the CellChat method is applied to assess the quantity, strength, and signaling pathway changes between the two cell subpopulations. Finally, at the spatial transcriptome level, the spatial distribution patterns of specific cell subtypes in different risk groups are identified, further revealing the spatial heterogeneity characteristics at the tissue microenvironment level.

[0080] In summary, this application systematically reveals the heterogeneity characteristics of GBM patients from multiple dimensions, including genomics, copy number, transcriptomics, proteomics, and immune microenvironment, through multi-omics integrated differential analysis under high- and low-risk groups.

[0081] In some embodiments, during the inter-omics correlation analysis in step S4, the correlation between copy number variation and transcriptome expression, as well as the correlation between transcriptome expression and proteome expression, are calculated to screen out intersection genes with significant correlations and consistent trends as key genes, so as to ensure the biological consistency of the screened genes at different omics levels.

[0082] In some embodiments, step S5 involves introducing the intersection genes into a public database and screening for genes that are significantly highly expressed in GBM to obtain candidate genes. This includes introducing the intersection genes into the TCGA and CGGA databases and screening for genes that are significantly highly expressed in GBM based on the difference in transcriptomic expression between GBM and low-grade gliomas, in order to enhance the tumor specificity of the candidate genes.

[0083] In some embodiments, step S6 involves introducing candidate genes into multiple public databases, combining survival analysis and statistical indicators to ultimately identify potential therapeutic targets related to the high-risk group. This includes: introducing candidate genes into multiple public databases, combining survival analysis, calculating the P-value of the Log-rank test and the P-value of the univariate Cox regression, screening out genes whose survival analysis P-values ​​at multiple omics levels are all less than 0.05, and counting and ranking them according to statistical indicators to ultimately identify potential therapeutic targets related to the high-risk group.

[0084] Specifically, this application further proposes a method for screening potential therapeutic targets related to high-risk groups based on multimodal risk stratification. Specifically, this application first performs differential analysis at the copy number variation (CNV) level, transcriptome level, and proteome level based on high- and low-risk groups obtained from a multimodal fusion deep model, respectively, to screen for a set of genes significantly overexpressed in the high-risk group. The intersection of the differential analysis results across omics is then used to preliminarily screen candidate genes significantly associated with the high-risk group. Based on this, this application further performs inter-omics correlation analysis, specifically calculating the correlation between copy number variation and transcriptome expression, and the correlation between transcriptome expression and proteome expression, to screen for key genes with high cross-omics consistency (significant correlation and consistent trend), ensuring the biological consistency of the screened genes at different omics levels. Subsequently, this application introduces the aforementioned intersection genes into the TCGA and CGGA databases, and based on the difference in transcriptome expression between GBM and low-grade glioma (LGG), further screens for genes significantly overexpressed in GBM to enhance the tumor specificity of candidate genes. In addition, this application preferably inputs the above-mentioned genes into multiple public databases, and calculates the P-value of the Log-rank test and the P-value of the univariate Cox regression in combination with survival analysis. Genes with P-values ​​of less than 0.05 in survival analysis at multiple omics levels are screened out, and they are counted and ranked according to statistical indicators to finally identify potential therapeutic targets related to the high-risk group.

[0085] In summary, the multimodal fusion-based intelligent prognostic prediction method for glioma provided in this application integrates multi-omics data to construct a multimodal fusion deep model. This multimodal data includes genomic data, copy number variation data, transcriptomic data, proteomic data, and radiomic data. The constructed multimodal fusion deep model is used for risk stratification, resulting in high- and low-risk groups. Heterogeneity analysis is then performed on the data from each omics. This heterogeneity analysis includes mutation profile analysis, genomic data analysis, copy number analysis, transcriptomic analysis, proteomic analysis, immune microenvironment analysis, single-cell analysis, and spatial transcription analysis, thereby obtaining differential analysis results for each omics. Finally, the target screening and validation process includes multi-omics differential analysis, multi-omics correlation analysis, multi-omics prognostic analysis, multi-database analysis, in vitro cell validation, and patient-derived xenograft (PDX) model validation.

[0086] See Figure 2This application also provides a multimodal fusion intelligent prognostic prediction system for gliomas. This system is used to implement the multimodal fusion intelligent prognostic prediction method for gliomas as described in the above embodiments. The system includes: a multimodal data acquisition module 101, a model construction module 102, a heterogeneity analysis module 103, and a therapeutic target screening module 104 connected in sequence. The multimodal data acquisition module 101 is used to collect multimodal data and extract omics features respectively. The model construction module 102 is used to construct a multimodal fusion deep model based on deep learning algorithms and omics features. The heterogeneity analysis module 103 is used to obtain high- and low-risk groups based on the multimodal fusion deep model, respectively from genomic and copy number variation perspectives. Differential analysis was performed across the dimensions of transcriptomics, proteomics, and immune microenvironment to obtain the results of differential analysis of each omics. The therapeutic target screening module 104 was used to screen out a set of genes that were significantly highly expressed in the high-risk group based on the results of differential analysis of each omics, and to obtain the intersection of the results of differential analysis of each omics. Candidate genes that were significantly associated with the high-risk group were initially screened out, and inter-omics correlation analysis was performed to screen out intersection genes that were significantly correlated and had consistent trends. The intersection genes were introduced into a public database to screen out genes that were significantly highly expressed in GBM to obtain candidate genes. The candidate genes were introduced into multiple public databases respectively, and combined with survival analysis and statistical indicators, to finally identify potential therapeutic targets related to the high-risk group.

[0087] In summary, this application adopts a standardized acquisition and fusion system for multimodal data, and for the first time proposes a unified acquisition, preprocessing, and quality control process for MRI images (magnetic resonance imaging), WSI (whole tissue scan), WES (exome sequencing), RNA-seq (transcriptome sequencing), and proteomics data. By establishing a horizontally consistent and vertically unified data standardization system, the horizontal comparability and vertical continuity of multi-omics data are significantly improved, laying a solid foundation for subsequent high-quality modeling and deep multimodal fusion.

[0088] Furthermore, this application employs heterogeneity mechanism analysis and biological validation. After completing the risk stratification of the fusion model, this application further systematically analyzes multi-omics heterogeneity characteristics, covering multiple dimensions such as gene mutation spectrum, functional pathway activity, immune microenvironment composition, single-cell transcriptome characteristics, and spatial heterogeneity changes. Through multi-level and multi-scale heterogeneity analysis, this application comprehensively validates the rationality and stability of different risk groupings from a biological mechanism perspective, greatly enhancing the interpretability and clinical translational value of the model.

[0089] Furthermore, this application proposes a screening and functional validation method for potential therapeutic targets. It innovatively integrates multi-omics data from multiple public databases to conduct cross-omics joint differential analysis, association analysis, and survival prognostic analysis, systematically screening potential therapeutic targets closely related to the high-risk group. Functional validation of the screened targets is achieved by combining in vitro cell experiments, mouse orthotopic tumor models, and patient-derived xenograft (PDX) models, successfully constructing a complete technical closed loop from risk prediction to precision treatment intervention, significantly enhancing the clinical application potential of the multimodal survival model.

[0090] Compared with existing technologies, the multimodal fusion-based intelligent prognostic prediction method and system for gliomas provided in this application have the following advantages: Compared with existing GBM risk prediction technologies based on simple fusion of single-omics or dual-modal approaches, this application achieves a significant breakthrough in the depth and breadth of data integration. By standardizing the acquisition and fusion of five modalities of data—MRI, WSI, WES, RNA-seq, and proteomics—this application can systematically characterize tumor heterogeneity at multiple levels, including imaging, tissue, and molecular aspects. This results in more accurate and stable risk stratification, with stronger generalization ability, significantly outperforming existing single- or dual-modal methods.

[0091] Regarding the fusion strategy, this application constructs a multimodal fusion deep model to obtain high- and low-risk groups, significantly improving the model's survival prediction performance and robustness. Based on project research data, the models in this application outperform existing similar methods, demonstrating the significant superiority of the fusion strategy.

[0092] In terms of biological validation, this application breaks through the limitation of traditional risk stratification models lacking interpretability. Through systematic multi-omics biological validation, it clearly reveals the essential differences between different risk groups in terms of gene mutation, pathway activity, immune microenvironment, etc., which greatly improves the credibility and clinical applicability of the model.

[0093] In terms of clinical translational applications, this application identified a new potential therapeutic target, EIF3I, based on risk grouping, and verified its intervention effect through in vitro and in vivo models, thus establishing a complete chain from prediction to intervention and providing a new scientific basis for personalized precision treatment.

[0094] Based on the above technical solutions, this application provides a multimodal fusion-based intelligent prognostic prediction method and system for gliomas. The method includes the following steps: First, multimodal data is collected, and omics features are extracted respectively. Then, a multimodal fusion deep model is constructed based on deep learning algorithms and omics features. Next, based on the multimodal fusion deep model, high- and low-risk groups are obtained, and differential analysis is performed from the dimensions of genome, copy number variation, transcriptome, proteome, and immune microenvironment to obtain the results of differential analysis of each omics. Then, based on the results of differential analysis of each omics, a set of genes that are significantly highly expressed in the high-risk group is screened, and the intersection of the results of differential analysis of each omics is obtained. Candidate genes that are significantly related to the high-risk group are initially screened, and inter-omics correlation analysis is performed to screen the intersection genes that are significantly correlated and have consistent trends. Next, the intersection genes are introduced into a public database to screen the genes that are significantly highly expressed in GBM to obtain candidate genes. Finally, the candidate genes are introduced into multiple public databases, and combined with survival analysis and statistical indicators, potential therapeutic targets related to the high-risk group are finally identified.

[0095] This application aims to overcome the shortcomings of existing multimodal modeling techniques in terms of fusion depth, data standardization, interpretability and validation, and clinical translational applications. It proposes a systematic, complete, and scalable new method for GBM risk stratification and target discovery based on multimodal data fusion, achieving an organic unity between accurate risk prediction and personalized treatment intervention. By constructing a multimodal fusion depth model, combined with heterogeneity mechanism analysis and target screening validation, this application comprehensively improves the accuracy, interpretability, and clinical translational potential of GBM patient stratification. This application significantly outperforms existing technologies in multiple dimensions, including data integration depth, model performance improvement, biological mechanism explanation, and clinical translational potential, demonstrating significant technological advancement and application promotion value.

[0096] Those skilled in the art will understand that the above-described embodiments are specific examples of implementing this application, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of this application. Any person skilled in the art can make their own modifications and alterations without departing from the spirit and scope of this application; therefore, the scope of protection of this application should be determined by the scope defined in the claims.

Claims

1. A multimodal fusion-based intelligent prognostic prediction method for gliomas, characterized in that, Includes the following steps: Collect multimodal data and extract omics features respectively; Based on deep learning algorithms and the aforementioned omics features, a multimodal fusion deep model is constructed; Based on the aforementioned multimodal fusion deep model, high- and low-risk groups were obtained, and differential analysis was performed from the dimensions of genome, copy number variation, transcriptome, proteome, and immune microenvironment to obtain the results of differential analysis of each omics. Based on the results of differential analysis of different omics, a set of genes that are significantly overexpressed in the high-risk group was screened out, and the intersection of the results of differential analysis of different omics was obtained. Candidate genes that are significantly related to the high-risk group were initially screened out, and inter-omics correlation analysis was performed to screen out the intersection genes with significant correlation and consistent trends. The intersecting genes were introduced into a public database, and genes that were significantly highly expressed in GBM were screened to obtain candidate genes. The candidate genes were introduced into multiple public databases, and by combining survival analysis and statistical indicators, potential therapeutic targets related to the high-risk group were finally identified.

2. The multimodal fusion-based intelligent prognostic prediction method for gliomas according to claim 1, characterized in that, Collect multimodal data and extract omics features separately, including: Acquire multimodal data, standardize the multimodal data, and extract the omics features corresponding to each omics from the standardized multimodal data; The multimodal data includes magnetic resonance imaging, whole tissue section scanning, exome sequencing, transcriptome sequencing, and proteome data; the omics features include radiomics features, pathomics features, whole exome sequencing data, transcriptome sequencing data, and proteome features.

3. The multimodal fusion-based intelligent prognostic prediction method for glioma according to claim 1, characterized in that, Based on deep learning algorithms and the aforementioned omics features, a multimodal fusion deep model is constructed, including: Multimodal data fusion is performed on the omics features corresponding to each omics, and a multimodal fusion hierarchical model is constructed based on deep learning algorithms.

4. The multimodal fusion-based intelligent prognostic prediction method for gliomas according to claim 1, characterized in that, Differential analyses were performed from the dimensions of genome, copy number variation, transcriptome, proteome and immune microenvironment to reveal biological differences among patients in different risk groups; At the genomic level, based on the Wilcoxon rank-sum test, the differences in gene mutation types, gene variant types, and single base substitution types among patients in different risk groups were compared.

5. The multimodal fusion-based intelligent prognostic prediction method for gliomas according to claim 4, characterized in that, At the copy number variation level, the limma method was used to identify chromosomal copy number variation segments that showed significant differences in different risk groups; At the transcriptomic and proteomic levels, genes and proteins with significantly different expression levels in high- and low-risk samples were screened based on the Wilcoxon rank-sum test. The screening criteria were FDR < 0.05 and absolute log2FC value greater than 1. Differentially expressed genes and proteins were then identified. ORA analysis was performed on the selected differentially expressed genes and proteins, and functional clustering and visualization of the enriched pathways were conducted.

6. The multimodal fusion-based intelligent prognostic prediction method for gliomas according to claim 4, characterized in that, At the level of immune microenvironment feature analysis, immune-related pathways involved in GSEA analysis were screened out, and their activation status differences were classified and visualized. Then, based on the ESTIMATE algorithm, the infiltration degree of stromal cells and immune cells in tumor tissue was estimated, and the differences between immune scores and stromal scores between the two risk groups were analyzed.

7. The multimodal fusion-based intelligent prognostic prediction method for gliomas according to claim 1, characterized in that, When conducting inter-omics correlation analysis, the correlation between copy number variation and transcriptome expression, as well as the correlation between transcriptome expression and proteome expression, is calculated. Intersecting genes with significant correlations and consistent trends are selected as key genes to ensure the biological consistency of the selected genes at different omics levels.

8. The multimodal fusion-based intelligent prognostic prediction method for gliomas according to claim 1, characterized in that, The intersecting genes were introduced into a public database, and genes that were significantly highly expressed in GBM were screened to obtain candidate genes, including: The intersecting genes were introduced into the TCGA and CGGA databases. Based on the difference in transcriptomic expression between GBM and low-grade glioma, genes that are significantly highly expressed in GBM were screened to enhance the tumor specificity of candidate genes.

9. The multimodal fusion-based intelligent prognostic prediction method for gliomas according to claim 1, characterized in that, The candidate genes were introduced into multiple public databases, and by combining survival analysis and statistical indicators, potential therapeutic targets associated with the high-risk group were finally identified, including: The candidate genes were introduced into multiple public databases, and survival analysis was performed. The P-values ​​of the Log-rank test and the univariate Cox regression were calculated to screen out genes with survival analysis P-values ​​less than 0.05 at multiple omics levels. Genes were counted and ranked according to statistical indicators to finally identify potential therapeutic targets related to the high-risk group.

10. A multimodal fusion intelligent prognostic prediction system for gliomas, the system being used to implement the multimodal fusion intelligent prognostic prediction method for gliomas as described in any one of claims 1 to 9, characterized in that, include: The system consists of a multimodal data acquisition module, a model building module, a heterogeneity analysis module, and a therapeutic target screening module, which are connected sequentially. The multimodal data acquisition module is used to collect multimodal data and extract omics features respectively; The model building module is used to construct a multimodal fusion deep model based on the deep learning algorithm and the omics features; The heterogeneity analysis module is used to obtain high- and low-risk groups based on the multimodal fusion depth model, and to perform differential analysis from the dimensions of genome, copy number variation, transcriptome, proteome and immune microenvironment to obtain the differential analysis results of each omics. The therapeutic target screening module is used to screen a set of genes that are significantly overexpressed in the high-risk group based on the results of differential analysis of various omics, and to obtain the intersection of the results of differential analysis of various omics. Candidate genes that are significantly associated with the high-risk group are initially screened, and inter-omics correlation analysis is performed to screen for intersection genes with significant correlation and consistent trends. These intersection genes are then introduced into a public database to screen for genes that are significantly overexpressed in GBM, thus obtaining candidate genes. These candidate genes are then introduced into multiple public databases, and combined with survival analysis and statistical indicators, potential therapeutic targets associated with the high-risk group are finally identified.

Images