High-throughput gene sequencing data analysis system based on deep learning
A high-throughput gene sequencing data analysis system based on deep learning has solved the problem of dynamic modeling of tumor clones in existing technologies, achieved deep fusion and precise quantification of multi-source data, and improved the accuracy of tumor clone evolution risk assessment and the reliability of clinical decision-making.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-16
- Publication Date
- 2026-05-05
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In existing technologies, high-throughput gene sequencing data analysis lacks a unified time benchmark and a systematic longitudinal integration method, making it difficult to dynamically model the growth, decline, and competition processes of tumor clones under treatment selection pressure and immune selection pressure. Tissue biopsy is highly invasive and difficult to repeat at high frequencies, and single-point sampling is easily affected by the spatial heterogeneity within the tumor, failing to fully represent the diversity of tumor clones.
A high-throughput gene sequencing data analysis system based on deep learning was adopted, including a data fusion and expression module, a variant feature generation module, a clonal immune dynamics analysis module, and a future risk inference module. By constructing a unified tensor representation, generating deep feature vectors, and integrating multi-source data, the clonal phylogenetic structure was reconstructed, the clonal growth rate and immune clearance intensity were estimated, and the risk of immune escape was inferred.
It achieves deep fusion of multi-source and multi-batch sequencing data, accurately quantifies the dynamic process of tumor clonal evolution, improves the ability to assess the risk of clonal evolution driven by key variants, provides an early warning window for clinical intervention, and enhances the interpretability and reliability of decision-making.
Smart Images
Figure CN121983126A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bioinformatics technology, and more specifically, to a high-throughput gene sequencing data analysis system based on deep learning. Background Technology
[0002] With the development of high-throughput gene sequencing and immunotherapy technologies such as immune checkpoint inhibitors, exome sequencing, whole-genome sequencing, circulating tumor DNA sequencing, and immune receptor rearrangement sequencing have been increasingly applied to medication decisions and efficacy evaluation in patients with advanced cancer. In clinical practice, deep sequencing results from a single or a few tumor biopsy samples are often relied upon, combined with mutation profiles, tumor mutation burden, and clonal structure inferences, to formulate or adjust immunotherapy regimens that last for several months.
[0003] However, tissue biopsies are highly invasive, risky, and difficult to repeat frequently. Single-point sampling is also susceptible to the spatial heterogeneity within the tumor, making it difficult to comprehensively represent the diversity of tumor clones. Existing analytical workflows often treat sequencing data from different time points as static, independent snapshots, lacking a unified time benchmark and a systematic longitudinal integration method. They also fail to dynamically model the growth, decline, and competition processes of tumor clones under treatment and immune selection pressures. Summary of the Invention
[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a high-throughput gene sequencing data analysis system based on deep learning to address the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides the following technical solution: A high-throughput gene sequencing data analysis system based on deep learning includes the following modules: The data fusion and representation module is used for quality control and unified coordinate mapping of multi-source sequencing reads to construct a unified tensor representation; The mutation feature generation module, connected to the data fusion and expression module, is used to generate deep feature vectors of candidate mutation sites based on the unified tensor representation, form a comprehensive feature vector by fusing external information, and calculate the comprehensive mutation risk index using a machine learning model. The clonal immunodynamics analysis module, connected to the variant feature generation module, is used to integrate the comprehensive risk index of variants at multiple time points with immune receptor rearrangement sequencing data, reconstruct the clonal phylogenetic structure, and estimate the intrinsic growth rate, interclonal competition coefficient, and immune clearance intensity coefficient of each clone, thereby obtaining the clonal abundance time series. The future risk reasoning module connects to the clonal immune dynamics analysis module. It is used to construct clonal evolution feature vectors based on clonal abundance time series, clonal intrinsic growth rate, interclonal competition coefficient, immune clearance intensity coefficient, and comprehensive mutation risk index. The clonal evolution feature vectors are then mapped to immune escape risk indexes to achieve the ranking and classification of high-risk clones.
[0006] In a preferred embodiment, when constructing a unified tensor representation, the data fusion expression module first normalizes or standardizes each feature channel, and then sets missing markers for genomic regions with low coverage and unmeasurable regions.
[0007] In a preferred embodiment, the mutation feature generation module generates deep feature vectors through a convolutional neural network.
[0008] In a preferred embodiment, the convolutional neural network includes at least two convolutional layers and a global pooling layer.
[0009] In a preferred embodiment, the machine learning model used to calculate the comprehensive risk index of mutations in the mutation feature generation module is a gradient boosting tree model.
[0010] In a preferred embodiment, the clonal immune dynamics analysis module fits and estimates the clonal abundance time series by constructing a dynamic model, which simultaneously couples the clonal autonomous proliferation capacity characterized by the intrinsic growth rate of each clone, the interclonal interaction characterized by the interclonal competition coefficient, and the immune system clearance effect characterized by the immune clearance intensity coefficient.
[0011] In a preferred embodiment, the parameters of the kinetic model are estimated using a nonlinear least squares method, and biological rationality constraints are imposed during the solution process.
[0012] In a preferred embodiment, the clonal evolution feature vector constructed by the future risk reasoning module includes trajectory features extracted from the clonal abundance time series, the intrinsic growth rate of clones, the immune clearance intensity coefficient, the inter-clonal competition coefficient, and derived features obtained by statistically summarizing the comprehensive variation risk index of all variations within the clone.
[0013] In a preferred embodiment, the future risk reasoning module uses a support vector machine model to map clonal evolution feature vectors to an immune escape risk index.
[0014] In a preferred embodiment, the system establishes a complete traceability chain between modules, enabling the final output immune escape risk index to be traced back to the underlying original sequencing evidence and related external information.
[0015] The technical effects and advantages of this invention are as follows: The high-throughput gene sequencing data analysis system based on deep learning provided by this invention effectively overcomes the heterogeneity of multi-source and multi-batch sequencing data by constructing a unified tensor representation, and realizes the deep fusion of multimodal and cross-time series data. By integrating gene sequencing and immune receptor rearrangement data and constructing a dynamic model that couples clonal proliferation, competition and immune clearance, it achieves a precise quantitative description of the dynamic process of tumor clonal evolution, breaking through the limitation of existing technologies that treat multi-time point data as static snapshots.
[0016] Building upon this foundation, the system innovatively integrates deep sequence features extracted through deep learning with multi-source external biological knowledge, significantly enhancing its ability to accurately and prospectively assess key variants and their driving clonal evolutionary risks, providing a valuable early warning window for clinical intervention. Simultaneously, the system design preserves the complete chain of evidence tracing back from the top-level risk index to the underlying evidence, and can clearly indicate key risk factors, greatly enhancing the interpretability of decision-making and the reliability of clinical applications. Attached Figure Description
[0017] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a schematic diagram of the structure of the high-throughput gene sequencing data analysis system based on deep learning of the present invention; Figure 2 This is a flowchart of the clonal immunodynamic analysis module of the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] Example 1: The high-throughput gene sequencing data analysis system based on deep learning of the present invention, such as... Figure 1 As shown, it includes the following modules: The data fusion and expression module is used to request the sequencing platform to provide raw read files during the data access stage, and at the same time retrieve information such as the batch number, sampling time, tissue source and key experimental parameters corresponding to each sample from the experimental management system. It attaches a unique sample identifier and time tag to each read through internal mapping rules.
[0020] After completing the basic annotation, this module initiates a preset quality assessment and cleaning process, checking the reads based on indicators such as base quality, linker residue, proportion of repeated reads, length distribution of inserted fragments, and overall coverage uniformity. For reads with consistently low quality, suspected contamination, or abnormally high repeatability, the module performs end truncation, local error correction, or complete removal according to established strategies, ensuring that the noise level of the data entering downstream analysis is controllable.
[0021] Provided quality control is passed, the data fusion and expression module uniformly aligns reads to a designated reference genome, forming structurally consistent alignment records. Each record clearly annotates the corresponding genomic coordinates, matching quality, strand orientation information, and events such as mismatches, insertions, and deletions, while retaining structural information such as pairing relationships between paired reads and insertion fragment lengths. For modes involving splicing or rearrangement, such as RNA sequencing and immune receptor rearrangement sequencing, the module uses a specific internal mapping strategy during the alignment stage to process cross-exon alignments, V(D)J rearrangements, and other cases, transforming all types of alignment results to the same reference coordinate system. This ensures the comparability of spatial locations between different sequencing strategies and platforms, avoiding systematic biases caused by inconsistencies in coordinate systems.
[0022] After obtaining alignment results in a standardized format, the data fusion and expression module divides the reference sequence according to pre-defined genomic windows or position indices, and summarizes multidimensional sequencing features within each window. For the position with sample number i and reference genomic position or window index j, the module extracts coverage statistics, base symbols and their probability distribution, matching quality distribution, insert length and read orientation patterns, and neighborhood sequence features from the alignment record, arranging them in a fixed order into several feature channels to form a three-dimensional tensor. Here, index i identifies the sample source, j identifies the specific location or window on the reference genome, and k corresponds to different types of feature channels, such as coverage depth, percentage of variant-supported reads, average alignment quality, paired read orientation combination, or local GC content. To mitigate systematic differences between different batches and sequencing depths, this module constructs... Normalization or standardization is performed on each channel, and missing markers are set for regions with low coverage or unmeasurable areas, so that the tensors of different samples are comparable on a numerical scale.
[0023] The mutation feature generation module receives the unified tensor representation output by the data fusion expression module. Based on this, sequencing evidence near a specific location and external knowledge about that location are organized into variant-level features for easy use in subsequent analysis. A set of suspicious sites obtained from a preliminary screening is given in advance. This module processes these sites one by one, rather than scanning indiscriminately across the entire genome, in order to reduce unnecessary computation and highlight signals related to variants.
[0024] When processing a candidate position j, the mutation feature generation module first extracts a local segment centered at that position from the 3D tensor, and expands it to both sides using a window with a half-width of w to obtain a local tensor. This process can be represented as ; In the formula, w is the window half-width, representing how many bases or windows are taken upstream and downstream from the candidate position. For example, in one embodiment, the window half-width w is set to 10; interval : indicates a continuous interval including candidate positions; : indicates that no filtering is performed on the feature channel dimension, and all channels are retained as is.
[0025] The variant feature generation module will The input is fed into a pre-trained deep feature extraction network.
[0026] After transformation by the deep feature extraction network, the system maps the local tensor into a low-dimensional vector. , recorded as .
[0027] in, The deep feature extraction network is a nonlinear mapping that compresses a high-dimensional tensor into a fixed-length vector. This represents the structured signal representation of candidate position j on sample i.
[0028] In one specific embodiment, the deep feature extraction network It is implemented using a lightweight convolutional neural network structure, the specific components of which are as follows: The first convolutional layer uses 64 filters of size 3×3 with a stride of 1, employs Same padding, and is followed by a ReLU activation function and a batch normalization layer.
[0029] The second convolutional layer uses 128 filters of size 3×3 with a stride of 1, employs Same padding, and is followed by a ReLU activation function and a batch normalization layer.
[0030] Global average pooling layer: compresses the feature maps output by the aforementioned convolutional layers into a 128-dimensional feature vector. .
[0031] To enable the feature extraction network to distinguish between variants of different pathogenicity levels, a sample set containing a large number of annotated variant sites is selected as training data during the training phase. This sample set includes both clinically confirmed benign and pathogenic variants, and each variant carries a corresponding pathogenicity category label. Network training can employ supervised learning with cross-entropy as the objective function. This means the cross-entropy between the model's output class probability distribution and the true label is used as the loss metric, and this loss is reduced through iterative iteration. For optimization, the Adam algorithm can be used for parameter updates. The initial learning rate is set to, for example, 0.001, the batch size to, for example, 128, and each iteration through all training samples is considered a training cycle. The total number of training cycles is set to, for example, 50. Under these settings, through multiple rounds of iterative optimization, the network parameters gradually converge, resulting in a more efficient and efficient network. The obtained feature vector It can effectively distinguish different categories of variation and provide stable and reusable representations for downstream risk assessment models.
[0032] Sequencing evidence alone is often insufficient to determine the importance of a variant. Therefore, the variant feature generation module also incorporates information related to the variant from external tools and databases. For each candidate location, this module can read several score values and classification labels from functional prediction tools, such as whether it has a destructive impact on protein structure, whether it may alter splicing sites, and the evolutionary conservation of the site. Simultaneously, it can obtain the allele frequency of the variant in different populations and whether it belongs to an existing common polymorphic site from population databases, and extract whether it is repeatedly observed in tumor samples, recorded as a driver variant, or associated with a specific pathway from disease or tumor-related databases. After importing, the aforementioned external information undergoes missing data imputation and numerical standardization before being combined with local deep feature vectors. The features are concatenated in a fixed order to form a comprehensive feature vector. Inside the module, Each dimension has a clear explanation, including its origin, range of values, and correspondence with biological processes, which facilitates subsequent tracking and interpretation.
[0033] After the comprehensive feature construction is completed, the variant feature generation module compresses the high-dimensional feature vector into a single variant comprehensive risk index using a risk calculation function. This index measures the functional impact of the variant in the current sample. This process can be written as follows: ; in, The aforementioned comprehensive feature vector contains local sequencing evidence and multi-source external information; Let S(i) be the covariate vector for sample i, used to represent the overall clinical background information at the patient level. The covariates may include at least one or more of age, stage, previous treatment regimen, and baseline immune status. S(⋅) is the risk calculation function. In one specific embodiment, the risk calculation function S(⋅) is implemented using a gradient boosting tree model. This model consists of 300 decision trees, each with a maximum depth of 5 and a learning rate of 0.05. The training objective of the model is a binary classification task, and the training data is the comprehensive feature vector. And its corresponding pathogenicity label, confirmed by public databases and clinical guidelines. The model outputs a comprehensive risk index of variants. This is the probability value predicted by the model that this variant causes disease. is the comprehensive risk index of the variant at candidate position j in sample i. A higher value indicates a greater potential impact of the variant on disease development, tumor clonal dominance, or immune escape. During the training phase, the system adjusts the parameters of S(⋅) using existing data on pathogenic variants, benign variants, and clinical outcomes to ensure the calculated... It has a reasonable consistency with real clinical risks; during the operational phase, the parameters remain fixed, and candidate variants in new samples are scored using a uniform standard.
[0034] The clonal immunodynamics analysis module starts with the variant level numerical representation output by the variant feature generation module. It combines the temporal information obtained from longitudinal sequencing with the results of immune receptor sequencing to characterize the compositional structure of different clones within the tumor and their changing trends over time using relatively intuitive parameter forms. During runtime, the module reads the variant risk index and allele frequency time series corresponding to each candidate variant. Simultaneously, it classifies and ranks these variants based on co-occurrence patterns, mutual exclusion relationships, and existing functional annotations. By constraining the compatibility of frequency changes over time and organizing the logical relationships of which variants must appear first and which can only appear under specific conditions, this module provides a clonal phylogenetic tree that can interpret the observed data. Mutations on the same branch are grouped into a single clonal unit, labeled with clone number q. Combining the allele frequency trajectory and copy number status of the variants contained in the clone, the module converts the frequency information at the variant level into a clonal abundance time series, denoted as q. Here, t is a time variable representing the relative proportion of the clone in the overall tumor cell population at different sampling time points. This provides a well-structured time-series input for subsequent kinetic analysis.
[0035] After obtaining multiple clonal abundance time series, the clonal immunodynamic analysis module treats these curves as observations and uses a simplified first-order kinetic model to describe the amplification and competition behavior of each clone under limited resource conditions. This module uses an ordinary differential equation to represent the abundance change of clone q as follows: .
[0036] In the formula, Represents the relative abundance of clone q at time t; This represents the relative abundance of tumor clone p in the tumor cell population at time point t; The intrinsic growth rate of clone q; is the interclonal competition coefficient, used to measure the inhibitory strength of clone p on clone q; p in the summation represents the indices of other clones besides q. This module uses the least squares method to minimize the time series of clonal abundance predicted by the model compared to the time series obtained from actual sequencing. The mean squared error between the two values is used to estimate a set of growth rates that are consistent with the data. Competition coefficient matrix between clones In the specific implementation, the Levenberg-Marquardt algorithm, a nonlinear least squares optimizer, is used to solve for the parameters.
[0037] Based solely on resource competition, the clonal immune dynamics analysis module further incorporates the clearance function of the immune system explicitly into the aforementioned model. To this end, the system first utilizes immune receptor rearrangement sequencing data to construct a matching set of immune receptor clones for each tumor clone. Then, based on affinity prediction and changes in receptor abundance, it obtains the effective immune receptor abundance time series for clone q. Subsequently, a decay term describing immune clearance is added to the aforementioned equation, making the clonal kinetic equation rewrite as: .
[0038] In this formula, This represents the abundance of effective immune receptors against clone q at time t, which can be understood as the strength of the immune cell population capable of recognizing or killing that clone. is the immune clearance intensity coefficient, used to characterize the actual clearance efficiency of clone q per unit of immune receptor abundance. A larger value indicates a more significant decrease in the abundance of that clone once recognized by the immune system. Similar to models containing only growth and competition terms, this module performs a joint fitting of clonal abundance curves and immune receptor abundance curves to... The relevant parameters were estimated so that the fitted equation could explain the overall changes in the tumor cell population and also reflect the phenomenon of sudden contraction of some clones when immune pressure increased.
[0039] Based on the results provided by the clonal immunodynamics analysis module, the future risk reasoning module offers quantitative judgments regarding what will happen in the coming period. It focuses on each identified tumor clone, and its core task is to estimate, given its historical evolutionary trajectory and current immune environment, whether the clone will continue to expand or even gain immune escape advantage within a preset future time window. To this end, the module reads the abundance time series output by the clonal immunodynamics analysis module. Internal growth rate Immune clearance intensity coefficient Information such as interclonal competition parameters and clonal phylogenetic structure is collected. Simultaneously, combined with the comprehensive mutation risk index output by the mutation feature generation module, relevant mutations within the same clone are analyzed to obtain a comprehensive profile of the clone in terms of functional risk, neoantigen burden, and immune-related pathway disturbances. When possible, background information such as treatment intervention time points, administration methods, lesion sites, and baseline immune status can be added to make subsequent reasoning more closely resemble actual clinical procedures.
[0040] Based on the above inputs, the future risk reasoning module constructs a clonal evolution feature vector for each clone q. During construction, it will be derived from the clonal abundance time series. Several representative time points were selected, and the abundance values at these time points were arranged chronologically to form a simplified historical trajectory, reflecting the upward, downward, or stable trend of the clone during previous follow-ups. Subsequently, the intrinsic growth rate was... and immune clearance intensity As an independent numerical feature, it quantitatively describes how fast the clone grows naturally without external intervention and how sensitive it is to clearance under prior immune stress. Correspondingly, the comprehensive risk index of each variant within the clone, obtained from the variant feature generation module, is summarized using methods such as maximum value, average value, count, and weighted summation to obtain several statistics reflecting functional risk and neoantigen load, which are also included in the calculation. For clones whose positions in the cloning tree are relatively clear, the relationships between them and their ancestor clones, child clones, and neighboring clones with strong competition can be added to the feature vector using a simplified encoding method. This makes the vector contain both the clone's own attributes and retain some contextual information.
[0041] With the comprehensive features already constructed, the future risk reasoning module uses a risk mapping function to compress the high-dimensional features into a single, easily interpretable risk index. This process can be formally written as... ; in, Let F(⋅) be the clonal evolution feature vector of the aforementioned clone q; F(⋅) is the risk mapping function, used to map multidimensional features to a real number; in a specific embodiment, the risk mapping function F(⋅) is implemented using a support vector machine model. This model uses a radial basis function as the kernel function, with the regularization parameter C set to 1.0 and the kernel function parameter gamma set to scale. The training data of the model is the clonal evolution feature vector. This includes a binary label indicating whether clinical recurrence or radiographic progression will occur within the next 12 months. The model outputs a risk index. The decision function value calculated for the model; This represents the risk index of immune escape for clone q within a preset future time period. Intuitively, The higher the value, the greater the likelihood that the clone will continue to expand or regain dominance in the future, assuming the current treatment strategy and host immune status remain unchanged; the calculated value... It can correspond to the actual disease outcome in a statistical sense.
[0042] During actual system operation, the future risk reasoning module will process all clones. The system sorts and classifies the values, marking clones that exceed a set high-risk threshold. This data can be aggregated at the patient level to provide an overall future risk assessment. For clones identified as high-risk, the system can also... Factors contributing significantly to risk are extracted, such as whether the individual carries a high-functioning risk variant, whether the neoantigen load is high, whether there was a brief contraction followed by a rapid rebound during previous treatment, or whether the individual holds a competitive advantage. These factors provide the basis for generating written explanatory materials or clinical decision support reports. The module always retains the data from... Tracing back to the clonal evolution feature vector Then, by tracing back to the output parameters of the clonal immunodynamic analysis module and the variant feature generation module, as well as the underlying sequencing data, each risk assessment can be traced back to the corresponding source of evidence, facilitating subsequent verification and supplementary explanation.
[0043] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0044] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.
[0045] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0046] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0047] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A high-throughput gene sequencing data analysis system based on deep learning, characterized in that, Includes the following modules: The data fusion and representation module is used for quality control and unified coordinate mapping of multi-source sequencing reads to construct a unified tensor representation; The mutation feature generation module, connected to the data fusion and expression module, is used to generate deep feature vectors of candidate mutation sites based on the unified tensor representation, form a comprehensive feature vector by fusing external information, and calculate the comprehensive mutation risk index using a machine learning model. The clonal immunodynamics analysis module, connected to the variant feature generation module, is used to integrate the comprehensive risk index of variants at multiple time points with immune receptor rearrangement sequencing data, reconstruct the clonal phylogenetic structure, and estimate the intrinsic growth rate, interclonal competition coefficient, and immune clearance intensity coefficient of each clone, thereby obtaining the clonal abundance time series. The future risk reasoning module connects to the clonal immune dynamics analysis module. It is used to construct clonal evolution feature vectors based on clonal abundance time series, clonal intrinsic growth rate, interclonal competition coefficient, immune clearance intensity coefficient, and comprehensive mutation risk index. The clonal evolution feature vectors are then mapped to immune escape risk indexes to achieve the ranking and classification of high-risk clones.
2. The high-throughput gene sequencing data analysis system based on deep learning according to claim 1, characterized in that: When constructing a unified tensor representation, the data fusion expression module first normalizes or standardizes each feature channel, and then sets missing markers for genomic regions with low coverage and those that are unmeasurable.
3. The high-throughput gene sequencing data analysis system based on deep learning according to claim 1, characterized in that: The mutation feature generation module generates deep feature vectors through a convolutional neural network.
4. The high-throughput gene sequencing data analysis system based on deep learning according to claim 3, characterized in that: A convolutional neural network contains at least two convolutional layers and one global pooling layer.
5. The high-throughput gene sequencing data analysis system based on deep learning according to claim 1, characterized in that: The machine learning model used to calculate the comprehensive risk index of mutations in the mutation feature generation module is a gradient boosting tree model.
6. The high-throughput gene sequencing data analysis system based on deep learning according to claim 1, characterized in that: The clonal immune dynamics analysis module fits and estimates the clonal abundance time series by constructing a dynamic model. The dynamic model simultaneously couples the clonal autonomous proliferation capacity characterized by the intrinsic growth rate of each clone, the interclonal interaction characterized by the interclonal competition coefficient, and the immune system clearance effect characterized by the immune clearance intensity coefficient.
7. The high-throughput gene sequencing data analysis system based on deep learning according to claim 6, characterized in that: The parameters of the kinetic model were estimated using the nonlinear least squares method, and biological rationality constraints were imposed during the solution process.
8. The high-throughput gene sequencing data analysis system based on deep learning according to claim 1, characterized in that: The clonal evolution feature vector constructed by the future risk reasoning module includes trajectory features extracted from the clonal abundance time series, the intrinsic growth rate of clones, the immune clearance intensity coefficient, the interclonal competition coefficient, and derived features obtained by statistically summarizing the comprehensive variation risk index of all variations within the clone.
9. The high-throughput gene sequencing data analysis system based on deep learning according to claim 1, characterized in that: In the future risk reasoning module, a support vector machine model is used to map clonal evolution feature vectors into an immune escape risk index.
10. The high-throughput gene sequencing data analysis system based on deep learning according to claim 1, characterized in that: The system establishes a complete traceability chain between its modules, enabling the final output immune escape risk index to be traced back to the underlying original sequencing evidence and related external information.