A cloud computing-based biological data deep analysis method and system and a storage medium

Abstract

The application relates to the technical field of cloud computing, solves the technical problems that in the prior art, false positive interaction leads to insufficient proportion of driving genes, the success rate of verification is low, the final target list contains a large number of false positives, the result accuracy is low, a large amount of time and cost are consumed to eliminate false positive targets, and in particular relates to a biological data deep analysis method and system based on cloud computing and a storage medium, the steps of the method are as follows: collecting multi-source biological data, performing distributed normalization and feature vector processing to generate multi-omics feature vectors; then obtaining a network result through graph splicing mining; finally, containerization scheduling analysis and verification update are performed to form an analysis feedback report; the application fuses multi-omics features through exclusive coding and attention, eliminates false positive edges through a Bayesian inference feedback graph network, and encapsulates and strengthens learning scheduling to accelerate analysis and support breakpoint recovery.

Description

Technical Field

[0001] This invention relates to the field of cloud computing technology, and in particular to a cloud-based method, system, and storage medium for deep analysis of biological data. Background Technology

[0002] Cloud computing is a service model that provides elastically scalable computing resources, such as servers, storage, and software, on demand via the Internet. Current technologies typically involve manually uploading raw data, such as gene sequencing data, to cloud storage, manually installing and configuring various analysis software environments on virtual machines, then running omics data processing and functional analysis programs sequentially, and finally downloading the results to the local machine for viewing.

[0003] Existing technologies suffer from false-positive interactions, leading to insufficient driver gene representation, low validation success rates, and a large number of false positives in the final target list, resulting in low accuracy. Eliminating false-positive targets is time-consuming and costly. In practice, during a multi-omics study of gastric cancer, the mean edge probability of a key sub-network was low. The posterior probability of the interaction between proto-oncogenes and growth factors was right at the boundary. This interaction is a known phosphorylation-dependent relationship, but the multi-omics data did not contain modification information and there was no external knowledge injection. The system mistakenly deleted this edge, causing the sub-network to break at a critical point. Downstream molecular docking only screened isolated proteins for drugs, ignoring the co-binding pocket of the complex. As a result, the selected compounds were ineffective in cell experiments. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a cloud-based method, system, and storage medium for deep analysis of biological data. It solves the technical problems in existing technologies, such as insufficient proportion of driver genes due to false positive interactions, low verification success rate, and a large number of false positives in the final target list, resulting in low accuracy of results. Eliminating false positive targets consumes a lot of time and cost.

[0005] To solve the above technical problems, the present invention provides the following technical solution: The steps of the method are: collecting original multi-source biological data of the target organism, and obtaining a preprocessed biological dataset by performing distributed normalization processing on the original multi-source biological data; Autoencoder feature fusion is performed on the preprocessed biological dataset to generate a fused multi-omics feature vector. Distributed feature vector processing is then performed on the preprocessed biological dataset and the fused multi-omics feature vector to obtain the first multi-omics feature vector. The graph splicing probability matrix is ​​generated from the first multi-omics feature vector to obtain the edge probability matrix. Based on the edge probability matrix, the biological network mining results are obtained through feedback mining. The process of processing the edge probability matrix through feedback mining includes: processing the edge probability matrix and the initial molecular graph through feedback iteration to obtain the enhanced node embedding and the first edge probability matrix. Key sub-confidence is evaluated based on the first-side probability matrix and enhanced node embedding, generating key sub-networks and network confidence; The network confidence, the first side probability matrix, the enhanced node embedding, and the first multi-omics feature vector are subjected to secondary external feedback processing to obtain the biological network mining results. Based on the results of biological network mining, a containerized scheduling decision is obtained through processing to obtain task scheduling decisions. Based on the task scheduling decisions, an automated state analysis is performed to generate automated analysis results. An analysis feedback report is obtained by verifying and updating the feedback based on the automated analysis results.

[0006] Preferably, the original multi-source biological data is processed through distributed normalization, including: generating a unified data format in the cloud through distributed processing of the original multi-source biological data; Cleaned biological data is obtained by performing cleaning and noise reduction based on a unified cloud data format. The cleaned biological data were normalized to obtain a preprocessed biological dataset.

[0007] Preferably, the autoencoder feature fusion based on the preprocessed biological dataset includes: processing the preprocessed biological dataset through an autoencoder to generate omics-specific coding vectors; Multi-head self-attention processing is performed based on omics-specific encoding vectors to generate an attention weight matrix; Cross-group feature fusion is performed on the attention weight matrix to generate a fused multi-omics feature vector.

[0008] Preferably, distributed feature vector processing is performed on the preprocessed biological dataset and the fused multi-omics feature vector, including: obtaining fusion model parameters by distributed training processing based on the fused multi-omics feature vector and the preprocessed biological dataset; Feature importance is evaluated based on fusion model parameters to generate omics importance weights; The importance weights of omics and the fused multi-omics feature vectors are processed into multi-omics feature vectors to obtain the first multi-omics feature vector.

[0009] Preferably, the graph stitching probability matrix generation of the first multi-omics feature vector includes: obtaining an initial molecular graph by graph stitching based on the first multi-omics feature vector; The node embedding representation is obtained by performing forward propagation of edge weights based on the initial molecular graph. The probability matrix of the initial molecular graph and node embedding representation is estimated to generate the edge probability matrix.

[0010] Preferably, the decision-making process based on the biological network mining results is containerized, including: generating a containerized analysis pipeline based on the biological network mining results through containerized orchestration. Based on containerized analysis pipelines, cloud resource awareness processing is used to obtain the resource status matrix; The resource state matrix and containerized analysis pipeline are processed for scheduling decisions to obtain task scheduling decisions.

[0011] Preferably, the automated state analysis based on task scheduling decisions includes: performing state feedback processing based on task scheduling decisions and containerized analysis pipelines to obtain an execution state vector; The recovery scheduling instruction is obtained by performing backtracking checks based on the execution state vector. Automated analysis of execution state vectors and recovery scheduling instructions is performed to generate automated analysis results.

[0012] Preferably, the verification update feedback process based on the automated analysis results includes: obtaining a verification report by performing bio-verification processing based on the automated analysis results; Based on the results of automated analysis, performance evaluation is conducted, and resource performance reports are generated. The verification report and resource performance report are updated and fed back to obtain an analysis feedback report.

[0013] The technical solution also provides a cloud computing-based biological data deep analysis system, which includes: a preprocessing module that collects raw multi-source biological data of the target organism and obtains a preprocessed biological dataset through distributed normalization processing based on the raw multi-source biological data; The omics feature module performs autoencoder feature fusion based on the preprocessed biological dataset to generate a fused multi-omics feature vector. Distributed feature vector processing is then performed on the preprocessed biological dataset and the fused multi-omics feature vector to obtain the first multi-omics feature vector. The biological mining module generates a graph stitching probability matrix from the first multi-omics feature vector to obtain an edge probability matrix. Based on the edge probability matrix, the biological network mining results are obtained through feedback mining. The automatic analysis module processes the results of biological network mining through containerized scheduling decision-making to obtain task scheduling decisions, performs automated state analysis based on task scheduling decisions, and generates automated analysis results. The analysis and feedback module generates an analysis and feedback report based on the automated analysis results through verification and update feedback processing.

[0014] According to one aspect of some embodiments of this disclosure, a computer-readable storage medium is provided having computer program instructions stored thereon that, when executed, implement the steps of the cloud computing-based deep analysis method for biological data mentioned above.

[0015] By employing the above technical solutions, the present invention provides a method, system, and storage medium for deep analysis of biological data based on cloud computing, which has at least the following beneficial effects: 1. This invention addresses the issue of feature loss caused by the unified encoding of multi-omics data due to their different characteristics through dedicated encoders for each omics and multi-head cross-omics weight learning. On one hand, genomic data captures long-distance sequence associations, proteomics data extracts peak patterns layer by layer, and metabolomics data uses an autoencoder to compress the concentration matrix unsupervised, ensuring that each omics encoding vector retains the complete information of its own data structure. On the other hand, multi-head self-attention uses three encoding vectors as the weights for inter-omics interactions, compressing the fusion features to 256 dimensions, quantifying the contribution of each omics to phenotypic prediction, and allowing subsequent analysis to prioritize the omics type with the highest contribution, thus improving the interpretability of multi-omics fusion features.

[0016] 2. This invention solves the problem of false positive interactions interfering with node embedding in static edge weights of a database by iteratively coordinating the Bayesian inference edge probability matrix with message passing in a graph neural network. First, Beta-Bernoulli conjugate inference uses the inner product of node embeddings as the interaction score to calculate the existence probability of each potential edge, retaining only high-confidence edges, thus significantly increasing the proportion of known driving genes in key subnetworks. Second, after replacing the static weights in the database with the edge probability matrix, it is re-injected into GraphSAGE for secondary message passing, allowing the node embedding iterations to gradually get rid of the interference of initial noisy edges, improving the embedding quality. Finally, cloud-based literature mining and retrieval of known associations correct the Bayesian prior parameters, enabling low-confidence subnetworks to obtain external evidence enhancement, thus improving network reliability.

[0017] 3. This invention solves the problems of multi-software environment conflicts and low resource utilization of static scheduling by using Docker to encapsulate analysis tools and dynamic scheduling via a deep Q network. On the one hand, each analysis tool and its dependent libraries are encapsulated as independent images, and a general workflow language defines the directed acyclic graph dependencies between steps, eliminating library version conflicts in cross-platform deployments and enabling the analysis pipeline to be used immediately in different cloud environments. On the other hand, the deep Q network outputs the optimal pairing of tasks and nodes based on the available CPU cores, memory, number of GPUs, and task queue length of each node in the cluster. This allows GPU-intensive tasks to be automatically assigned to GPU nodes and memory-intensive tasks to large memory nodes, significantly shortening the overall completion time compared to static scheduling. When a task fails, it distinguishes between software errors and insufficient resources for separate handling, improving recovery efficiency in abnormal scenarios.

[0018] 4. This invention calculates the validation rate through cross-validation with literature and clinical variant databases, thereby automatically quantifying the biological reliability of the analysis results. It also uses the validation rate threshold to trigger incremental learning and performance indicators to drive the adjustment of reward function weights. Together, these enable the optimization of data preprocessing parameters, incremental training of the fusion model, and adaptive adjustment of scheduling weights, ultimately improving the continuous self-evolution capability of cloud-based biological data analysis. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a cloud computing-based deep analysis method for biological data according to the present invention; Figure 2 This is a structural block diagram of a cloud computing-based biological data deep analysis system according to the present invention. Figure 3 This is a comparison chart of network confidence scores for a cloud-based deep analysis method for biological data according to the present invention. Figure 4 This is a comprehensive statistical chart of a cloud computing-based deep analysis method for biological data according to the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0021] Example 1: Due to the insufficient proportion of driver genes caused by false positive interactions in existing technologies, coupled with low validation success rates and a lack of confidence assessment and knowledge correction, the final target list contains a large number of false positives, resulting in low accuracy. Eliminating false positive targets consumes significant time and resources. Please refer to... Figure 1 This embodiment provides a cloud-based method for in-depth analysis of biological data, which can improve the success rate of verification, reduce false positive data, significantly reduce costs and improve efficiency. The method includes the following steps: S1. Collect raw multi-source biological data of the target organism, and obtain a preprocessed biological dataset by distributed normalization processing based on the raw multi-source biological data. Existing technologies have long single-machine parsing time for multi-format data and poor coding consistency. Quality filtering with fixed computing power cannot be elastically scaled, resulting in low processing efficiency. Multi-omics numerical features are directly stored, and the problem of small numerical markers being submerged due to the failure to eliminate the difference in units is not eliminated. To solve the above problems, the specific implementation steps are as follows: S11. Based on the original multi-source biological data, a unified cloud data format is generated through distributed processing. In this step, the first step is: the cloud application interface gateway receives the original multi-source biological data and writes it into a distributed message queue for caching; the second step is: the format parser pulls various biological data files from the distributed message queue and extracts data identifiers, data types, patient identifiers, and timestamps; the third step is: the unified cloud data format = {data identifier, data type, patient identifier, timestamp, standardized binary large object payload}. The distributed message queue uses Apache Kafka, whose partitioning mechanism supports high-throughput concurrent access of multi-source heterogeneous data such as genome sequencing files, proteomics mass spectrometry files, metabolomics files, and electronic medical record messages, smoothing out peaks and valleys to avoid cloud entry point congestion caused by sudden data flows. The format parser calls the sequencing file parser, mass spectrometry file parser, metabolomics file parser, and electronic medical record parser according to the data type, converting the raw binary data into a unified binary large object format for storage, eliminating the differences in file format and encoding methods of different data sources. The raw multi-source biological data includes: genome sequencing files (FASTQ format), proteomics mass spectrometry files (RAW format), metabolomics mass spectrometry files (mzML format), and electronic medical record messages (HL7 format). The unified data format in the cloud includes five fields: data identifier, data type (values ​​are genomics, proteomics, metabolomics, and electronic medical record), patient identifier, timestamp, and standardized binary large object payload.

[0022] S12. Cleaning and denoising are performed based on the unified cloud data format to obtain cleaned biological data. The first step in this process is as follows: For sequencing data of genomics type in the unified cloud data format, the base quality value is calculated as -10 × the logarithm of the base error probability at base 10. If the quality value of any base is less than 20, the sequencing fragment is discarded. The second step is: For mass spectrometry data of proteomics or metabolomics type, if the signal-to-noise ratio is less than 3, the peak is discarded. The third step is: Using a distributed computing framework, data with a quality value greater than 20 and a signal-to-noise ratio greater than 3 are retained as cleaned biological data. A base quality value of 20 corresponds to a base error probability of 0.01. Sequencing fragments with a base interpretation accuracy of less than 99% are below this threshold, and retaining them will lead to an increase in the false positive rate of downstream variant detection. The mass spectrometry signal-to-noise ratio threshold of 3 is a commonly used minimum standard for identifying peptides and metabolites in proteomics and metabolomics. Spectral peaks with a signal-to-noise ratio below 3 are drowned out by instrument noise and are difficult to distinguish from the real signal. The distributed computing framework Spark splits the filtering task into multiple partitions and executes them in parallel on multiple nodes. The elastic resource allocation dynamically increases or decreases the number of computing nodes according to the amount of data, so that omics data of different scales can be cleaned in minutes.

[0023] S13. Perform data normalization on the cleaned biological data to obtain a preprocessed biological dataset. The first step in this process is to compare the minimum and maximum values ​​of numerical features such as gene expression levels and metabolite concentrations in the cleaned biological data across all samples. The second step is to calculate the normalized value as follows: Normalized value = (Original value - Minimum feature value) / (Maximum feature value - Minimum feature value). The third step is to hash-shard the normalized data according to chromosome number or sample identifier, store it in a distributed file system, set the block size to a preset value, and include the cloud storage path and metadata index in the dataset object. Max-min normalization maps gene expression levels (ranging from 0 to 90,000) and metabolite concentrations (ranging from 0.01 to 5,000 μmol / L) across different scales to a uniform range of 0 to 1. This eliminates the bias effect of numerical scale differences on subsequent gradient updates in deep learning models, ensuring a fair distribution of attention weights among genomics, proteomics, and metabolomics features. Hash sharding uses chromosome numbers or sample identifiers as hash keys to ensure that gene expression data from the same chromosome or multi-omics data from the same sample is stored in the same shard, reducing cross-node data read latency. A block size of 128MB is the default configuration for the distributed file system, achieving a balance between metadata storage overhead and parallel read efficiency.

[0024] This invention achieves efficient access, standardized cleaning, and numerical scale unification of multi-source heterogeneous biological data through a progressive design of distributed message queue unified format parsing, parallel quality filtering, and minimum-maximum value normalized hash sharding storage. This enables subsequent omics-specific encoders to be trained on clean and normalized data, ensuring that the input features of subsequent molecular network construction are not biased by dimensional differences. From the source of data preprocessing, this invention guarantees the accuracy and efficiency of cloud-based biological data deep analysis throughout the entire process.

[0025] S2. Based on the preprocessed biological dataset, autoencoder feature fusion is performed to generate a fused multi-omics feature vector. Distributed feature vector processing is then performed on the preprocessed biological dataset and the fused multi-omics feature vector to obtain the first multi-omics feature vector. Existing technologies use a unified encoder to ignore differences in data characteristics, resulting in insufficient feature extraction. Fixed weight concatenation or summation cannot automatically learn the importance of omics interactions, leading to the dilution of key signals. Single-machine training cannot be flexibly scaled, resulting in excessively long analysis cycles for large-scale cohorts. The fusion model lacks quantitative evaluation of omics contributions, making the results lack biological interpretability. To solve the above problems, the specific implementation steps are as follows: S21. Based on the preprocessed biological dataset, an autoencoder is used to generate omics-specific encoding vectors. The first step in this process is to distribute the preprocessed biological dataset to different graphics processor nodes according to three types: genomics, proteomics, and metabolomics. The second step involves encoding genomics data into genomics vectors using a pre-trained DNA language model, encoding proteomics data into proteomics vectors using a one-dimensional convolutional neural network, and compressing metabolomics data into metabolomics vectors using a fully connected autoencoder. The third step is to define the omics-specific encoding vector as: {genomics vector, proteomics vector, metabolomics vector}. Different encoders were used for the three types of omics data to suit their respective data characteristics. Genomics data consists of long-sequence base sequences, and a pre-trained DNA language model was used to capture long-distance regulatory element associations in the sequence using its attention mechanism, outputting a 768-dimensional vector. Proteomics mass spectrometry data consists of one-dimensional peak sequences, and a three-layer one-dimensional convolutional neural network was used to extract local peak patterns and global peptide distribution features layer by layer, outputting a 512-dimensional vector. Metabolomics data consists of numerical concentration matrices, and a fully connected autoencoder was used. The encoder compressed the data to a 256-dimensional bottleneck layer, and then the decoder reconstructed the original values ​​for unsupervised pre-training. The bottleneck layer output was used as the encoding vector. The three encoders were executed in parallel on independent GPU nodes, avoiding the time accumulation of single-node serial encoding. The genomics data were encoded using a pre-trained DNABert model, which consists of a stacked 12-layer Transformer encoder with 768 hidden layers and a maximum sequence length of 512. It was trained on a human reference genome using a masked language model loss for approximately 500,000 to 1,000,000 steps, terminating when the validation set perplexity did not decrease for five consecutive epochs, and outputting a 768-dimensional vector. The proteomics data were encoded using a three-layer one-dimensional convolutional neural network, with kernel sizes of 7, 5, and 3 layers, and channel numbers of 32, 64, and 128 respectively. It was trained on a mass spectrometry dataset using a classification loss for 10... The algorithm runs for 0 to 200 cycles, terminating when the verification loss does not decrease for 10 consecutive cycles, and outputs a 512-dimensional vector. Metabolomics data is compressed using a fully connected autoencoder. The encoder maps metabolite features layer by layer to a 256-dimensional bottleneck layer. It is trained on the concentration matrix with the reconstruction mean square error for 150 to 300 cycles, terminating when the verification loss does not decrease for 15 consecutive cycles, and the 256-dimensional vector of the bottleneck layer is taken as the encoded output. The three encoding methods are executed in parallel on independent GPU nodes, and each requires only one forward propagation during online inference. Pre-trained weights or fixed training parameters make the encoding results reproducible.

[0026] S22. Perform multi-head self-attention processing based on omics-specific coding vectors to generate an attention weight matrix. The first step involves concatenating the genomics vector, proteomics vector, and metabolomics vector into a matrix with 3 rows and columns equal to the dimension of the attention head. The second step involves mapping this matrix into a query matrix, a key matrix, and a value matrix through three different linear transformations. The third step is: Attention weight matrix = Flexible maximum function ((Query matrix × Transpose of key matrix) / Square root of key matrix column dimension) × Value matrix. The multi-head self-attention mechanism takes the encoding vectors of the three omics as input. The dot product of the query matrix and the key matrix measures the interaction strength between any two omics. Dividing by the square root of the column dimension of the key matrix is ​​to prevent the gradient vanishing due to an excessively large dot product value. The column dimension of the key matrix is ​​768, which is equal to the dimension of the genomics vector. When concatenating the matrices, the proteomics vector and the metabolomics vector are aligned to the same dimension through linear projection. The flexible maximum function normalizes the interaction strength to a weight between 0 and 1, and then multiplies it with the value matrix to achieve adaptive weighted fusion of each omics feature in the context. The output attention weight matrix has a dimension of 3 multiplied by 768, which preserves the complete weighted representation of each omics.

[0027] S23. Perform cross-group feature fusion on the attention weight matrix to generate a fused multi-omics feature vector. In this step, the first step is to add the three attention output vectors of genomics, proteomics and metabolomics in the attention weight matrix element by element to obtain the additive vector. The second step is to fuse the multi-omics feature vector = linear rectified activation function (fully connected weight matrix × additive vector + fully connected bias vector). Additive fusion directly adds the vectors of the three omics after attention weighting element by element. This operation assumes that the features of each omics are linearly additable in the fusion space, which is computationally efficient and has a short gradient propagation path, avoiding dimensional expansion after concatenation. The fully connected layer maps the fused vector to 256 dimensions. The dimension of the fully connected weight matrix is ​​256 times the dimension of the attention head, and the dimension of the fully connected bias vector is 256. This output dimension is consistent with the dimension of the metabolomics encoding vector, so that the fused features are on the same order of magnitude as the single-omics encoding, which makes it easier for the graph neural network in the parent step three to construct node features with a unified dimension. The linear rectified activation function introduces a nonlinear transformation, enabling the fused features to express complex cross-omics interactions.

[0028] S24. Based on the fusion of multi-omics feature vectors and preprocessed biological datasets, distributed training is used to obtain the fusion model parameters. In this step, the first step is to calculate the local gradient on each graphics processor node; the second step is to synchronize the local gradients of all nodes through the full reduction algorithm, and the global gradient = sum of local gradients / total number of nodes; the third step is to calculate the model parameters as the original model parameters - learning rate × global gradient, and output the fusion model parameters after training. The Ring-AllReduce synchronization algorithm organizes the GPU nodes into a ring topology. Each node communicates only with its neighboring nodes to transmit gradient slices. The communication overhead is independent of the number of nodes, and the gradient synchronization time remains stable as the number of nodes increases. The learning rate is between 0.0001 and 0.01, with the initial value usually set to 0.001 and decayed by a factor of 0.5 when the validation loss stagnates. This ensures that the model parameter updates can converge quickly without oscillating around the optimal solution. The global gradient is the arithmetic mean of the local gradients of all nodes, ensuring that the model parameters of each node are strictly consistent after each batch of training. The Ring-AllReduce synchronization algorithm is a commonly used communication algorithm, which will not be elaborated here.

[0029] S25. Based on the fusion model parameters, evaluate the importance of features and generate omics importance weights. The first step in this process is to extract the attention weight matrix from the fusion model parameters and decompose it into attention output vectors for each of the three dimensions: genomics, proteomics, and metabolomics. The second step is to calculate the importance weight of genomics as the mean of the genomics attention vector, the importance weight of proteomics as the mean of the proteomics attention vector, and the importance weight of metabolomics as the mean of the metabolomics attention vector, all obtained using the mean formula. The third step is to calculate the omics importance weight as: {genomics importance weight, proteomics importance weight, metabolomics importance weight}. The attention weight matrix is ​​calculated in S22 using a multi-head self-attention mechanism. Each element reflects the degree of attention a omics feature receives in the fusion context. The average attention of each omics is obtained by averaging the element-wise attention output vectors of each omics, with values ​​ranging from 0 to 1. The sum of the importance weights of the three omics is 1. When the importance weight of a certain omics is significantly higher than that of other omics, it indicates that the omics contributes the most to the current phenotypic prediction, providing a quantitative basis for prioritizing the omics in subsequent analyses. In S26, the omics importance weights are appended as metadata to the fused multi-omics feature vector and output together, providing prior weight information of each omics feature for the construction of the molecular network graph in S31, and also marking the degree of contribution of each omics to the disease-related module in the final result of S36.

[0030] S26. Perform multi-omics feature vector processing on the omics importance weights and the fused multi-omics feature vector to obtain the first multi-omics feature vector; the first step in this step is to extract the 256-dimensional fusion feature vector body from the fused multi-omics feature vector, and extract the genomics importance weight, proteomics importance weight, and metabolomics importance weight from the omics importance weights; the second step is: the first multi-omics feature vector = {fusion feature vector body, omics importance weight metadata}; The omics importance weights are derived from the statistical averaging of the attention weight matrix by omics dimension in S25, with the sum of the three weights being 1. This weight is appended to the fused multi-omics feature vector as metadata, ensuring that when constructing the molecular network graph in downstream S31, the node features not only contain a 256-dimensional fusion vector but also carry information on the contribution ratio of each omics to the node's representation. When outputting the biological network mining results in S36, the contribution of each omics to key sub-networks can be labeled, enhancing the biological interpretability of the mining results. The first multi-omics feature vector includes a 256-dimensional real-valued vector and omics importance weight metadata. The real-valued vector is generated by S23 through additive fusion and fully connected mapping, fusing attention weighting information from genomics, proteomics, and metabolomics. The omics importance weight metadata consists of three scalar values: genomics importance weight, proteomics importance weight, and metabolomics importance weight, obtained by averaging the attention weight matrix by omics dimension in S25, with the sum of the three being 1.

[0031] This invention addresses the shortcomings of existing technologies where genomics, proteomics, and metabolomics data are typically analyzed independently and then simply spliced ​​together. Each omics uses the same encoding method, ignoring its individual data characteristics, leading to insufficient feature extraction. In practice, a cancer cohort used a single fully connected network to encode the three omics data uniformly. This resulted in the loss of long-range regulatory information in the genomic base sequence after being flattened into vectors, and the local patterns of proteomic mass spectrometry peaks were globally averaged. The fused features failed to reflect the hierarchical relationship between gene regulation and protein expression, resulting in lower disease classification accuracy. This invention employs a pre-trained DNABert model for genomics to capture long-range regulatory element associations in sequences, a three-layer one-dimensional convolutional neural network for proteomics to extract local peak patterns and global distributions layer by layer, and a fully connected autoencoder for metabolomics to unsupervisedly compress the concentration matrix. The three encoders run in parallel on independent GPU nodes, preserving the integrity of the intrinsic structure of each omics and enabling the self-attention mechanism to calculate cross-omics interaction weights on fully encoded features, thereby enhancing the quality of omics representation before feature fusion.

[0032] By employing a multi-head self-attention mechanism with each omics encoding vector as input, the interaction strength between any two omics is calculated through the dot product of the query matrix and the key matrix, and the gradient is stabilized. Additive fusion adds the attention outputs of the three omics element-wise and maps them to 256 dimensions through linear rectified activation. This not only automatically learns the task-related omics interaction weights but also quantifies the contribution of each omics by the mean of the attention weights, thus ensuring that the fused features are both compact and retain discriminative information. This solves the problem that existing technologies, which use direct splicing or fixed weighted summation, cannot automatically learn the importance of different omics features in the prediction of specific phenotypes, resulting in the dilution of key signals. In practice, in a metabolic disease study, metabolomics features play a decisive role in disease classification, but fixed-weight splicing fusion causes genomics and proteomics features to occupy most of the dimensions of the feature space, diluting the key differences in metabolomics in thousands of dimensions of spliced ​​vectors, and causing the attention of the classification model to be scattered by irrelevant features.

[0033] Specifically, the above-mentioned approach can solve the problem in practice that when training multi-omics fusion models on a single machine, the training sample size or the number of omics types increases, the models cannot be flexibly scaled and the synchronization efficiency of parameters at each node is low, resulting in excessively long training cycles. In practice, a fusion model of whole genome, proteomics and metabolomics for a certain cohort of 1,000 people requires about nine hours to train once on a single workstation, and optimization iterations of 100 times take more than a month. During this period, hyperparameters cannot be adjusted, and model optimization is severely limited. In a special case during training, due to insufficient video memory of a certain graphics processor node, the batch size was forced to decrease locally, the gradient estimation variance increased, and the model convergence curve oscillated violently. In this case, the attention weights trained in the end exhibited non-physical negative feedback between genomics and proteomics, causing the fusion features to lose biological interpretability. This invention organizes the graphics processor nodes into a ring topology using the Ring-AllReduce synchronization algorithm. Each node communicates only with its neighboring nodes. The global gradient is the arithmetic mean of the local gradients. The communication overhead is independent of the number of nodes. The learning rate is 0.001 and is halved when the validation loss stagnates. This not only reduces the training time of multi-omics models in large-scale queues from months to days, but also appends the mean of the attention weights after training as metadata of omics importance to the output, thereby ensuring the scalability of the model while retaining the interpretable contribution information of each omics.

[0034] S3. Generate a graph concatenation probability matrix from the first multi-omics feature vector to obtain an edge probability matrix. Based on the edge probability matrix, obtain the biological network mining results through feedback mining. Existing technologies directly use static weights from the database, resulting in insufficient proportion of driver genes due to false positive interactions and low verification success rate. Step-by-step execution leaves low-probability edge residues, causing a small number of irrelevant pathways. The lack of confidence assessment and knowledge correction results in a large number of false positives in the final target list, leading to low accuracy. Eliminating false positive targets consumes a lot of time and cost. To solve the above problems, the specific steps are as follows: S31. Based on the first multi-omics feature vector, an initial molecular graph is obtained through graph splicing. The first step in this process is to identify a list of significantly expressed genes or proteins from the first multi-omics feature vector. The second step is to map the list of molecules to graph nodes using a protein-protein interaction database and a gene regulation database as the framework. The third step is to assign edge weights equal to the combined score of the molecule pair extracted from the protein-protein interaction database divided by 1000. The fourth step is to assign initial node features equal to the feature components corresponding to the molecule in the first multi-omics feature vector, and then combine all nodes, edges, and node features to form the initial molecular graph. The comprehensive score of the protein interaction database ranges from 0 to 1000. After dividing by 1000, it is normalized to the 0-1 interval and directly used as the initial edge weights of the graph neural network. The gene regulation database provides the regulatory relationship between transcription factors and target genes. Together, they constitute the skeleton topology of the initial molecular graph. The node features are taken from the feature components corresponding to each molecule in the fusion multi-omics feature vector output from the mother step two, so that the nodes carry fusion information of genomics, proteomics and metabolomics. The protein interaction database is a public bioinformatics database that records the physical contact, functional association and regulatory relationship between proteins. The STRING database is one of the most widely used protein interaction databases. Each interaction record is accompanied by a comprehensive score, ranging from 0 to 1000. The higher the score, the stronger the evidence of interaction.

[0035] S32. Based on the initial molecular graph, perform forward propagation of edge weights to obtain the node embedding representation. In this step: First, for each node, calculate the normalized edge weight of all its neighboring nodes = the weight of that neighboring edge / the sum of the weights of all neighboring edges of that node; Second, weighted neighbor information aggregation = multiplying the embedding vector of each neighboring node's previous layer by its normalized edge weight and then summing the results; Third, current layer embedding = linear rectified activation function (current layer weight matrix × concatenating the node's previous layer embedding with the weighted neighbor information aggregation); Fourth, using the database score in the initial molecular graph as the initial edge weight, after two layers of propagation, the final embedding vectors of all nodes constitute the node embedding representation. The edge-weighted GraphSAGE forward propagation uses the normalized comprehensive score of the protein interaction database as the weight for aggregating neighbor information. The higher the database score, the greater the contribution of the interaction relationship to the node embedding. The two-layer propagation enables each node to aggregate molecular interaction information within its second-order neighborhood. The receptive field radius corresponds to the indirect association range of about 2 to 5 proteins in the biological network, which precisely covers the local module structure of protein complexes and signaling pathways. The linear rectified activation introduces a nonlinear transformation, enabling the embedding to express the superposition effect of multi-level regulatory relationships. The embedding dimension is consistent with the input feature dimension, both being 256 dimensions, which is consistent with the dimension of the fused multi-omics feature vector output by S2.

[0036] S33. Generate the edge probability matrix by estimating the probability matrix of the initial molecular graph and node embedding representation; In this step, the first step is: for each pair of unconnected nodes in the initial molecular graph, the interaction score = the inner product of the embedding vector of the first node and the transpose of the embedding vector of the second node; the second step is: set the prior to follow a beta distribution with shape parameters of initial alpha and initial beta, the posterior alpha = initial alpha + the number of known true positive edges between the pair of nodes, the posterior beta = initial beta + the number of known true negative edges between the pair of nodes; the third step is: edge existence probability = posterior alpha / (posterior alpha + posterior beta), and the edge existence probabilities of all unconnected node pairs constitute the edge probability matrix; Beta-Bernoulli conjugate inference uses the inner product of node embedding vectors as an interaction score to reflect the strength of the potential association between two molecules; the higher the interaction score, the stronger the prior bias. The initial alpha and initial beta are both set to 1.0, corresponding to the uninformed beta prior, that is, the prior mean of the edge existence probability is 0.5 when no known edge information is known, indicating complete uncertainty. The posterior parameter is updated by the statistics of true positive and true negative edges in the known database, so that the probability estimation simultaneously integrates the embedding similarity learned by the graph neural network and the empirical evidence from the biological knowledge base. This edge probability matrix is ​​used as a feedback weight in S34 to re-inject into the graph neural network to drive the secondary enhancement of node embedding, and key subnetworks are extracted by probability thresholding in S35.

[0037] S34. Based on the edge probability matrix and the initial molecular graph, the enhanced node embedding and the first edge probability matrix are obtained through feedback iteration. In this step, the first step is to use the existence probability of each edge in the edge probability matrix as the new edge weight, replacing the original database score weight in the initial molecular graph. The second step is to re-execute the graph neural network message passing based on edge weight weights to generate the updated node embedding. The third step is to re-execute the Beta-Bernoulli conjugate inference based on the updated node embedding to calculate the updated edge probability matrix. The fourth step is to calculate the change in node embedding. If the change is greater than the convergence threshold, return to the first step to continue iterating. If it is less than the convergence threshold or the preset number of iterations is reached, terminate the iteration and output the enhanced node embedding and the final edge probability matrix. The number of iterations is preset to 2 to achieve a balance between computational overhead and convergence performance. The convergence threshold is set to 0.01 to 0.05 to control the tolerance of the embedding vector variation. In the first iteration, the database score weights are used to generate the initial embedding and initial edge probabilities. In the second iteration, the edge probabilities estimated by Bayes replace the database score weights and the message passing is performed again. This allows the node embedding to absorb multi-omics feature information and prior validation evidence from the biological knowledge base at the same time. The edge probabilities and node embeddings reinforce each other in the iteration. This feedback mechanism eliminates the single dependence of the initial edge weights on the static database score. The graph neural network employs a two-layer edge-weighted GraphSAGE architecture. Each layer performs a nonlinear transformation by weighting and aggregating neighbors using a linear rectified activation function and concatenating them with its own features. The number of iterations is fixed at two and does not change with the input. The termination condition is that the change in the Frobenius norm of the node embedding matrix is ​​less than the convergence threshold of 0.01 to 0.05 or the preset number of iterations of two is reached. The number of iterations is set to two. The most significant change in the embedding vector is between the first iteration, which uses database weights, and the second iteration, which uses Bayesian estimated edge probabilities. After Bayesian probabilities correct a large number of false positive edges in the database, the second message passing can immediately correct noise interference, resulting in a qualitative improvement in embedding quality. The marginal benefits of the subsequent third and fourth iterations decrease sharply, and the embedding tends to stabilize.

[0038] S35. Based on the first edge probability matrix and enhanced node embedding, perform key sub-confidence evaluation to generate key sub-networks and network confidence. In this step, the first step is to traverse each edge in the final edge probability matrix ZJZ_bgl. If the edge has a probability GL_bc greater than the set probability threshold YZ_gl, retain the edge b; otherwise, discard it. The second step is to perform connected component analysis on the graph formed by the retained edges, extracting the connected subgraph containing the most known disease-related genes and the largest number of nodes as the key sub-network WL_gj. The third step is to calculate the mean edge probability Ju. Z_bgl = Sum of the edge existence probabilities GL_bc of all retained edges in the critical subnetwork WL_gj / Total number of edges N in the critical subnetwork; Step 4: Edge probability entropy SH_bgl = Sum of -(Edge existence probability GL_bc of each edge in the critical subnetwork WL_gj × its natural logarithm + (1 - edge existence probability GL_bc) × ln(1 - edge existence probability GL_bc)); Step 5: Network confidence = Mean edge probability JuZ_bgl × e^-(Edge probability entropy SH_bgl / Total number of edges in the critical subnetwork); A probability threshold of 0.9 is used to retain only high-confidence edges with a probability higher than 90% following Bayesian inference, effectively filtering out false positive interactions with low initial scores or those contradicting multi-omics characteristics in the protein interaction database. Connected component analysis decomposes the graph formed by high-confidence edges into unconnected modules, extracting the largest connected subgraph containing known disease-related genes to ensure consistency between the mining results and current biological knowledge. Network confidence consists of two parts: the mean edge probability reflects the average reliability of edges in the key subnetwork, and the edge probability entropy in the exponential term, after being normalized by the number of edges, measures the dispersion of each edge probability. When all edge probabilities are close to 1, the entropy approaches 0 and the confidence approaches 1; when the edge probabilities are uneven, the entropy increases and the confidence decreases.

[0039] Experimental data analysis: such as Figure 3 As shown, a multi-omics cancer cohort dataset was used as the test object. This dataset contains genomic, proteomic, and metabolomic data of 300 tumor samples. An initial molecular graph was constructed using the STRING and RegNetwork databases as the backbone. After processing through S32 to S34, the final edge probability matrix was obtained. Three methods were compared 30 times each: whole-network module extraction without probability screening, screening with a fixed threshold of 0.5 and no confidence assessment, and screening with a probability threshold of 0.9 plus network confidence assessment as proposed in this invention. The experiment showed that whole-network module extraction without probability screening included all edges in the initial molecular graph for analysis. The extracted largest connected subgraph contained 1847 nodes, but known cancer driver genes accounted for only 12% of them, and the false positive rate of enrichment analysis was as high as 32%. The fixed threshold of 0.5 screening method retained edges with an edge probability greater than 0.5, reduced the number of sub-network nodes to 623, and increased the proportion of known driver genes to 31%. However, due to the lack of confidence assessment, the quality of the sub-network could not be quantified, and some low-probability edges were still retained, resulting in 11% irrelevant pathways in the downstream functional enrichment results. The method of this invention retains only high-confidence edges with a probability threshold of 0.9, reducing the number of sub-network nodes to 207. The proportion of known cancer driver genes is significantly increased, and the calculated network confidence is 0.87. The average edge probability is 0.94, and the edge probability entropy is only 0.12, indicating that the retained edges not only have a high average probability but also good consistency. When 30 random low-probability noise edges are manually injected into the edge probability matrix, the no-probability screening method includes all noise edges, while the fixed threshold method retains 14 noise edges. The method of this invention retains only one noise edge that accidentally exceeds the threshold due to the probability threshold of 0.9, and the network confidence drops to 0.61, triggering the secondary external knowledge feedback correction in S36. This verifies that the probability threshold screening effectively filters out false positive interactions, and the confidence assessment effectively quantifies the reliability of the sub-network.

[0040] S36. Perform secondary external feedback processing on network confidence, first edge probability matrix, enhanced node embedding, and first multi-omics feature vector to obtain the biological network mining results. In this step, the first step is: if the network confidence is greater than or equal to the confidence threshold, the current key sub-network and the final edge probability matrix are directly used as the final result. The second step is: if the network confidence is less than the confidence threshold, secondary feedback is triggered, the cloud literature mining service is started to retrieve known disease-related literature evidence for nodes in the key sub-network, the literature support is converted into beta prior parameter correction amount and summed with the initial alpha to update the prior parameters, and then beta Bernoulli conjugate inference, probability weighted graph convolution iteration, key sub-network extraction and confidence evaluation are executed again in sequence to obtain the corrected key sub-network and network confidence. The third step is: biological network mining result = {sub-network is taken from key sub-network, confidence is taken from network confidence, key node list is taken from the top ten nodes in the key sub-network in terms of degree, association omics importance is taken from the omics importance weight metadata in the first multi-omics feature vector, and node embedding is taken from enhanced node embedding}. The confidence threshold is set between 0.6 and 0.8. A threshold below this indicates that the probability of a large number of edges in the key subnetwork is low or inconsistent, requiring external literature evidence to enhance the network's credibility. The cloud-based literature mining service retrieves known disease-related literature for nodes in the key subnetwork from databases such as PubMed and DisGeNET, converting the co-occurrence frequency or association significance in the literature into a beta prior parameter correction amount. The correction amount ranges from 0 to 5. After correction, the beta prior is updated to adjust the edge probability estimation towards the direction of literature evidence. After re-executing the complete inference chain from S33 to S35, the corrected key subnetwork integrates evidence from multiple omics features, known interaction databases, and the latest literature, thus improving the network's confidence. The list of key nodes is sorted based on node degree. The higher the degree, the more centrally the molecule is in the key subnetwork.

[0041] This invention addresses the problem of directly using static scores from a protein interaction database as edge weights. This leads to a large number of false positive interactions and missing edges in the database not being corrected, resulting in a network with significant noise. In practice, some protein interaction relationships in the STRING database originate from text mining co-references rather than physical experimental verification. Directly using these as initial edge weights in a graph neural network allows irrelevant proteins to interfere with neighbor embeddings during message passing, ultimately resulting in known disease-driving genes accounting for less than 15% of the key subnetwork. This invention uses Bayesian inference with the inner product of node embedding vectors as the interaction score. Through Beta-Bernoulli conjugate inference, it transforms known edge relationships in the database into posterior parameter updates, outputting the existence probability matrix for each potential edge. With a probability threshold of 0.9, only high-confidence edges are retained. This not only significantly increases the proportion of known driver genes but also allows the edge probability matrix, after replacing the static weights in the database, to be re-injected into the graph neural network for secondary message passing. This ensures that the interference from false positive edges on node embeddings is continuously weakened, enhancing the realism of the network structure.

[0042] By feeding the edge probability matrix obtained from Bayesian estimation as new edge weights into the graph neural network, re-execute message passing to generate updated node embeddings, and then recalculate edge probabilities based on the new embeddings, with the number of iterations preset to 2 and the convergence condition being that the change in the Frobenius norm is less than 0.01 to 0.05, this approach not only achieves synergistic optimization of edge probabilities and node embeddings, allowing the embeddings to gradually escape initial database noise during iteration, but also ensures that the final extracted key subnetworks are built on the embeddings enhanced by probability feedback. This guarantees the consistency and accuracy of the network mining results and solves the problem of isolated execution of graph neural network message passing and edge probability estimation, where node embeddings are generated only based on fixed initial edge weights and cannot utilize subsequent probability inference results to optimize embedding quality. In practice, traditional GraphSAGE generates embeddings all at once after the graph structure is determined. Even if downstream path analysis finds that some edges cannot exist under actual conditions, the embeddings cannot be corrected, leading to a small number of irrelevant paths in subsequent functional enrichment analysis.

[0043] Specifically, the above scheme can solve the problems of lack of quantitative evaluation of network confidence, equal weighting of low-quality and high-quality sub-networks, lack of external knowledge injection mechanism to correct when the overall network confidence is low, low mean probability of key sub-network edge, and direct use of the sub-network as the sole input for downstream drug target analysis. The posterior probability of the interaction edge between proto-oncogene and growth factor falls near the decision boundary. This interaction has been confirmed in the literature to be phosphorylation condition-dependent, and can only stably bind to growth factor when a specific phosphorylation site of proto-oncogene is activated. However, the lack of phosphorylation modification information in multi-omics data leads to ambiguity in Bayesian posterior probability. Without external knowledge injection, the system erroneously deletes this edge, the sub-network breaks at this key connection, and downstream molecular docking only screens drugs for isolated proteins without considering the co-binding pocket of the complex. The final screened compounds are ineffective in cell experiments. This invention processes network confidence using the mean of side probabilities, with a confidence threshold ranging from 0.6 to 0.8. When the confidence level falls below the threshold, a cloud-based literature mining service is activated to search the PubMed and DisGeNET databases. The literature support is then converted into a beta prior parameter correction amount and re-inferred. This not only provides targeted reinforcement of external knowledge to low-confidence subnetworks but also encapsulates and outputs the corrected list of key nodes and omics importance weights, thereby ensuring that the final biological network mining results are both data-driven and literature-verifiable.

[0044] S4. Based on the results of biological network mining, containerized scheduling decision processing is used to obtain task scheduling decisions. Based on the task scheduling decisions, automated state analysis is performed to generate automated analysis results. Existing technologies have dependencies and conflicts, resulting in time-consuming deployment and poor portability. Static scheduling strategies cannot dynamically match node resources according to task computation characteristics, resulting in long overall completion time. After a task fails, it is necessary to restart the entire process from the starting point, lacking breakpoint resumption and error type differentiation recovery mechanisms. Analysis results lack unified archiving and cost accounting. To solve the above problems, the specific steps are as follows: S41. Based on the results of biological network mining, a containerized analysis pipeline is generated through containerized orchestration. The first step involves extracting the node list and topology of key sub-networks from the biological network mining results, and automatically matching the analysis toolchain according to the biological function category of each node. The second step involves querying the gene function annotation database for each key node in the key sub-network to determine the biological pathways it participates in, generating a personalized analysis task queue. The third step is: the containerized analysis pipeline = {step sequence takes the personalized analysis task queue, input data is convoluted to extract key sub-network node features, output path}, where the personalized analysis task queue = {perform functional enrichment analysis on the key sub-network node list, perform molecular docking on the top few nodes with the highest degree in the key sub-network, and perform variant annotation on the nodes where the key sub-network intersects with the known disease mutation database}, and the step sequence is a directed acyclic graph. The key subnetwork is output from S3 after Bayesian inference and probability thresholding, and its node list is highly enriched with known disease-driving genes. The personalized analysis task queue determines the analysis priority based on the node degree. The top 5 nodes with the highest degree are prioritized for molecular docking with core hub proteins. Nodes that intersect with known disease mutation databases such as ClinVar or COSMIC are annotated with variants to assess their clinical relevance. The entire node list is subjected to gene ontology and KEGG pathway functional enrichment analysis to reveal the overall biological functional modules of the network. Each analysis tool and its dependent libraries are packaged as independent Docker images. A common workflow language defines the dependencies between steps, enabling tasks to be executed in the order of a directed acyclic graph. The input data volume is mounted with the node feature vectors of the key subnetwork, ensuring that the analysis pipeline can be ported and reproduced across cloud platforms.

[0045] S42. Based on the containerized analysis pipeline, cloud resource awareness processing is used to obtain the resource status matrix. The first step in this step is to collect the available CPU cores, available memory capacity, available GPUs, and current task queue length of each compute node in the cluster through the container orchestration application programming interface. The second step is to summarize the resource status matrix as follows: {set of available CPU cores, set of available memory capacity, set of available GPUs, set of task queue length}. The resource data of all nodes are summarized into this matrix. The resource state matrix provides a real-time cluster state view for the S43 reinforcement learning scheduler. The number of available CPU cores reflects the general computing power of a node, and tasks such as functional enrichment analysis and mutation annotation are preferentially scheduled to nodes with sufficient CPU power. The available memory capacity determines the number of analysis containers a single node can simultaneously handle; each bioinformatics container typically requires 2 to 8 gigabytes of memory. The number of available GPUs is a scarce resource for computationally intensive tasks such as molecular docking and deep learning inference, and is allocated only on GPU nodes. The task queue length reflects the current load backlog of a node; when the queue length exceeds 3, the node will no longer accept new tasks. The state matrix is ​​collected and updated every 10 to 30 seconds to balance scheduling real-time performance with application programming interface (API) call overhead.

[0046] S43. Perform scheduling decision processing on the resource state matrix and containerized analysis pipeline to obtain task scheduling decisions. The first step in this process is to concatenate the available CPU cores, available memory capacity, available GPUs, and task queue length of all nodes in the resource state matrix with the progress vector of the currently pending task in the containerized analysis pipeline to form a joint state vector ZX_lh. The second step is to use the joint state vector ZX_lh as input and perform forward propagation to output the Q-value of each optional action. The optional action is to assign a pending task to a certain computing node. The third step is to use the task scheduling decision RW_dc to randomly select an action with a probability of exploration rate G_ts and select the action with the largest Q-value with a probability of 1 - exploration rate G_ts, outputting the selected task and node pairing. The deep Q-network adopts a three-layer fully connected structure. The input layer dimension equals the resource state dimension plus the pipeline progress dimension, and the output layer dimension equals the number of tasks to be executed multiplied by the number of available nodes. The exploration rate is initially set to 1.0 and gradually decays to 0.05 during training to balance exploration and utilization. The decay step size is 0.001 to 0.005 per scheduling cycle. The reward function is designed as the negative of the total completion time and the weighted sum of resource idle costs. The Q-network updates parameters by minimizing the temporal difference error. The update target is the immediate reward plus a discount factor of 0.9 to 0.95 multiplied by the maximum Q value of the next state. The network terminates when the maximum number of training steps is reached. The maximum number of training steps is set to 10,000 to 30,000 steps. This invention sets it to 20,000 steps.

[0047] S44. Based on the task scheduling decision and containerized analysis pipeline, perform status feedback processing to obtain the execution status vector. The first step in this process is to pair the selected task with the node in the task scheduling decision, create a container instance in the container orchestration cluster, and mount the input data volume specified in the containerized analysis pipeline to the container. The second step is to capture the container's standard output stream and error output stream after it starts, recording the task identifier, execution status, duration, and exit code. The third step is to write the execution status vector = {task identifier, execution status (running, successful, or failed) enumeration value, duration, exit code} to the cloud database in real time. Container instances are created via the container orchestration application programming interface. Each container corresponds to an analysis task, and the task identifier is inherited from the personalized analysis task queue of S41. The input data volume mounts the node characteristics and intermediate result files of the key sub-network. The mounting method adopts a read-write consistency protocol to ensure that the data is ready when the container starts. An exit code of 0 indicates that the task has been successfully completed, while a non-zero code corresponds to different types of software errors or insufficient resources. The execution status is initially "running" and is updated to "success" or "failure" after the task is completed. This execution status vector provides real-time task status information for breakpoint resumption and error recovery in S45. When the status is "failure" and the exit code indicates a software error, a directed acyclic graph backtracking is triggered. When the exit code indicates insufficient resources, rescheduling to a larger resource node is triggered.

[0048] S45. Perform backtracking checks based on the execution state vector to obtain the recovery scheduling instruction. In this step: First, if the execution state vector indicates failure, check the exit code to distinguish the cause of failure. An exit code that is non-zero and not 137 indicates a software error; an exit code of 137 indicates insufficient resources. Second, if it is a software error, backtrack through the directed acyclic graph of the containerized analysis pipeline to locate the previous successful checkpoint of the current failed step. Read the intermediate output file of that checkpoint from the mounted storage volume. The recovery scheduling instruction is: {Retry step: take the current failed step; Last successful output path: take the checkpoint path}. Third, if it is insufficient resources, the recovery scheduling instruction is: {Retry step: take the current failed step; Target node: take a new node with more available CPU cores and more memory than the original node}. Each time a containerized analysis pipeline completes a step, a checkpoint file is written to the storage volume, saving all the output data for that step. The checkpoint retains the output of all successful steps for backtracking. Exit code 137 is the standard signal for system termination due to memory overrun in a container orchestration cluster, which clearly distinguishes between software anomalies and resource bottlenecks. Software errors are resolved by directly loading the intermediate file of the upstream checkpoint, allowing for breakpoint resumption without re-executing completed steps, thus shortening the recovery time from a full process rerun to a single-step retry.

[0049] S46. Perform automated result analysis on the execution state vector and recovery scheduling instructions to generate automated analysis results. The first step in this process is to iterate through the final states of all tasks in the execution state vector. If all are successful, the overall analysis state is set to success. If there are failures but the recovery scheduling instructions have been successfully retried, the overall analysis state is set to success; otherwise, it is set to failure. The second step is to write the final analysis output files, such as the functional enrichment analysis results, molecular docking results, and mutation annotation results, to cloud object storage and generate corresponding access paths. The third step is to calculate the automated analysis result as follows: Automated analysis result = {result file path, analysis state is set to success or failure, total time is the time difference between the start of the first task and the end of the last task, cost is calculated as: CPU core hour price × total core hour + GPU card hour price × total card hour + memory gigabyte hour price × total gigabyte hour}. The total execution time is obtained by summing the duration fields in the execution status vectors of each task in S44 and subtracting the waiting intervals for resuming from breakpoints. In cost calculation, the unit price for the CPU is 0.03 to 0.08 yuan per core per hour, and for this invention, it is 0.06 yuan per core per hour; the unit price for the GPU is 2 to 5 yuan per card per hour, and for this invention, it is 4 yuan per card per hour; the unit price for memory (Gbytes) is 0.005 to 0.01 yuan per Gbyte per hour, and for this invention, it is 0.008 yuan per Gbyte per hour. These are all reference ranges for standard public cloud billing. A successful analysis status indicates a key sub-network. All subsequent analyses have been fully executed and the results are available for download. In the event of failure, the path to the result file still retains the intermediate output of the successful steps for manual review. The automated analysis results include the following data: the cloud object storage access path for the final analysis output files such as functional enrichment analysis results, molecular docking results, and variant annotation results; the overall analysis status (success or failure); the total time taken from the start of the first task to the end of the last task; the total cloud resource cost calculated in terms of CPU core time, GPU card time, and memory gigabyte time; and a link to the complete analysis report.

[0050] This invention addresses the challenges of biological data analysis involving multiple software tools, conflicts in library versions and runtime environments, time-consuming and error-prone deployments across different computing platforms, and difficulties in migrating and reproducing analytical workflows. In practice, installation failures occur when installing statistical software packages for functional enrichment analysis and cheminformatics libraries for molecular docking on the same system due to shared library version conflicts. Researchers spend days debugging the environment but still cannot run two analytical steps simultaneously, ultimately resorting to manual step-by-step execution. This invention encapsulates each analytical software and its dependent libraries into independent Docker images, defines the directed acyclic graph dependencies between steps using a common workflow language, and generates standardized containerized analytical pipelines. This not only eliminates cross-platform deployment environment conflicts, allowing the analytical workflow to be installed and used immediately on any cloud platform, but also enables breakpoint resumption in case of task failure by backtracking the directed acyclic graph to the upstream checkpoint and restoring intermediate files from the storage volume, thus ensuring the portability and fault tolerance of the analytical workflow.

[0051] By using a reinforcement learning scheduler with the number of available CPU cores, memory, number of GPUs, and task queue length of cluster nodes as state inputs, and a deep Q-network outputting the Q-value of each task to be executed and assigning it to each node, the optimal action is selected using an ε-greedy strategy. This not only prioritizes GPU-intensive tasks to GPU nodes and memory-intensive tasks to large memory nodes, but also updates the resource status every 10 to 30 seconds, driving the scheduler to continuously perceive cluster changes. As a result, the overall completion time is significantly shortened compared to static scheduling. This solves the problem that multi-task analysis cannot dynamically adjust allocation based on the computational characteristics of tasks and the real-time load of nodes, leading to uneven resource utilization and extended overall completion time. In practice, molecular docking tasks require GPU acceleration, while functional enrichment analysis only requires a CPU. Static scheduling allocates GPU nodes to enrichment analysis, resulting in idle GPUs, while molecular docking tasks run slowly on CPU nodes for several hours, leading to a serious underutilization of overall cluster resources.

[0052] Specifically, the above solution addresses the practical problem of existing technologies failing to quickly recover from successful steps, resulting in a significant waste of computational resources and time. In practice, for example, in a genome-wide association study, the variant annotation step exited due to an input file format issue. Traditionally, this required a complete rerun starting from the raw data quality filter, taking eight hours. However, in this solution, only the annotation parameters need to be corrected before continuing from the annotation step. Furthermore, in a special case where a molecular docking task was terminated by the container orchestration system at 90% completion on a graphics processor node due to memory exceeding limits, the exit code was 137. If the system indicates insufficient resources, simply retrying the original node will exceed the limit again. It is necessary to identify the error type and reschedule to a node with more memory. This invention uses a breakpoint resumption mechanism to distinguish between software errors and insufficient resources by exit codes. When a software error occurs, the system backtracks the directed acyclic graph to the upstream successful checkpoint and recovers intermediate files from the storage volume. When resources are insufficient, the system reschedules to a node with at least twice the original allocated memory. This not only avoids redundant calculations of rerunning the entire process, but also writes the execution status and exit code of each task to the cloud database in real time, thereby ensuring that the entire analysis pipeline can be accurately located and quickly recovered when encountering various anomalies.

[0053] S5. Based on the automated analysis results, the analysis feedback report is obtained through verification and update feedback processing. The lack of systematic cross-validation of literature and databases in the existing technology leads to the failure to identify false positives. Without tracking and quantitative evaluation, cost efficiency cannot be optimized. The analysis process is executed in one direction and there is no channel to feed back the verification information and performance information to the front-end model. It is impossible to accumulate experience from historical cases to achieve continuous improvement. To solve the above problems, the specific implementation steps are as follows: S51. Based on the automated analysis results, a validation report is obtained through biovalidation. The first step in this process is to extract a list of key genes or metabolites from the automated analysis results and count the total number of entities in the list. The second step is to search for known associations between each entity in the list and the target disease in the literature database and the clinical variant database, and count the number of hits of known associations. The third step is to calculate the validation rate as (number of hits / total number) × 100%, and the validation report includes the validation rate and detailed annotation information of the hit entities. The literature database provides text mining results linking gene function to disease, while the clinical variant database records clinically verified pathogenic variants and their corresponding diseases. Together, they form an external knowledge verification benchmark. The number of hits is obtained through automated text mining and precise matching with the database. Literature hits require that the gene name and disease name appear together in the same sentence and meet the relevance threshold, while database hits require that the gene has a pathogenicity label in the variant annotation file. The verification rate ranges from 0% to 100%, with a higher value indicating a greater proportion of the automated analysis results supported by known literature or databases.

[0054] S52. Based on the automated analysis results, conduct a performance evaluation and generate a resource performance report. In this step, the first step is to extract the total time and total cost values ​​from the automated analysis results; the second step is to calculate the throughput as the amount of data analyzed and the total time consumed; the third step is to calculate the resource performance report as follows: the total computation time is the total time consumed, the total cost is the total cost consumed, the throughput is the throughput consumed, and optimization suggestions are included. Total time and total cost are directly inherited from the automated analysis results of S46, ensuring data consistency. The amount of data analyzed refers to the total amount of input data from the raw multi-source biological data input by S1 to the automated analysis results output by S4, measured in gigabytes. Throughput reflects the scale of data processed by the system per unit time; the larger the value, the higher the efficiency of cloud resource utilization. Optimization suggestions are automatically generated based on the execution time ratio of each analysis task. When the time taken by the molecular docking task exceeds 50% of the entire process, it is recommended to increase the number of graphics processor nodes. When the memory usage of the functional enrichment analysis is close to the node's memory limit, it is recommended to switch to a node with a larger memory specification. When the task queue length is continuously greater than 3 for multiple cycles, it is recommended to expand the cluster size.

[0055] S53. Update and process the verification report and resource efficiency report to obtain the analysis feedback report. The first step in this step is to compare the verification rate in the verification report with the preset threshold. The second step is to collect a list of key genes and metabolites from the error cases and feed them back to the distributed deep learning model in the parent step two. After loading the current fusion model parameters as initial weights, incremental training is performed on the encoder and attention layer. The third step is to extract the task time ratio and node utilization data from the resource efficiency report and increase the cost weight or time weight in the reinforcement learning reward function to change the scheduling preference. The fourth step is to generate an analysis feedback report = {verification rate, total computation time and total cost efficiency indicators, model update records, and suggestions for the next analysis}, which is then fed back to the sampling strategy in S1. The validation rate threshold is set between 60% and 80%. A value below this threshold indicates that many genes in the key sub-networks discovered in the parent step three are not supported by known literature or databases. It is necessary to incrementally train the fusion model using the current erroneous cases to improve the accuracy of subsequent analyses. During incremental training, the parameters of the fusion model are obtained from the parameter archive after training in sub-step 2.4 of the parent step two. The adjustment range of cost weight and time weight in the reinforcement learning reward function is set between 0.1 and 0.3. When the resource efficiency report shows that the cost is too high, the cost weight is increased to make the scheduler more inclined to select low-cost nodes. When the total time is too long, the time weight is increased to make the scheduler prioritize completion speed. The next analysis suggestion is automatically generated based on the validation rate and efficiency indicators, such as "If the validation rate is below 65%, it is recommended to expand the training set" or "If the GPU idle rate is above 40%, it is recommended to reduce the number of GPU nodes".

[0056] This invention calculates the validation rate through cross-validation with literature and clinical variant databases, thereby automatically quantifying the biological reliability of the analysis results. It also uses the validation rate threshold to trigger incremental learning and performance indicators to drive the adjustment of reward function weights. Together, these enable the optimization of data preprocessing parameters, incremental training of the fusion model, and adaptive adjustment of scheduling weights, ultimately improving the continuous self-evolution capability of the overall cloud-based biological data analysis.

[0057] Examples of experimental data: Figure 4 As shown, using a multi-omics cancer cohort dataset as the test object, which contains genomic, proteomic, and metabolomic data of 300 tumor samples, three end-to-end comparative experiments were conducted 30 times each using single-omics analysis with traditional static scheduling, simple multi-omics concatenation with fixed-edge weight network mining, and the full-process deep analysis method of this invention. The experiments showed that: the single-omics approach only uses gene expression data for analysis, and the proportion of known cancer driver genes in the key gene list is only 21%, with a total time of 4.7 hours and a total cost of about 320 yuan; the simple multi-omics concatenation approach directly concatenates the three omics features and constructs a network with static weights from the STRING database, increasing the proportion of driver genes to 37%, but the factor network contains a large number of false positive edges, resulting in about 15% irrelevant pathways in the downstream enrichment analysis, with a total time of 6.2 hours and a total cost of about 410 yuan; This invention employs DNA BERT to encode the genome, a one-dimensional convolutional neural network to encode the proteome, and a fully connected autoencoder to compress the metabolome. After multi-head self-attention fusion, the proportion of driving genes is increased to 68% through Bayesian inference of the edge probability matrix and iterative collaboration with a graph neural network. After screening with a probability threshold of 0.9, the number of sub-network nodes is reduced from 1847 to 207, achieving a network confidence of 0.87. Furthermore, dynamic scheduling using a deep Q-network reduces the total processing time to 3.1 hours and the total cost to approximately 195 yuan. Simultaneously, when the validation rate is below 65%, incremental learning is triggered to feed back into the fusion model, gradually improving the accuracy of subsequent analyses. The stepwise increase in the proportion of driving genes from 21% to 37% and then to 68% verifies the synergistic effect of multi-omics fusion and Bayesian network mining. The reduction in total processing time from 4.7 hours to 3.1 hours verifies the efficiency advantage of reinforcement learning scheduling over traditional static scheduling. This demonstrates that the invention's full-process closed-loop deep analysis method effectively achieves end-to-end optimization from multi-source data to a trusted network and then to efficient scheduling.

[0058] Example 2: Due to the insufficient proportion of driver genes caused by false positive interactions in existing technologies, low validation success rates, and a lack of confidence assessment and knowledge correction, the final target list contains a large number of false positives, resulting in low accuracy. Eliminating false positive targets consumes a significant amount of time and cost. Please refer to [link / reference needed]. Figure 2The diagram shown is a structural block diagram of a cloud computing-based biological data deep analysis system provided in this embodiment. The system includes a preprocessing module, an omics feature module, a biological mining module, an automatic analysis module, and an analysis feedback module. The preprocessing module collects raw multi-source biological data of the target organism and obtains a preprocessed biological dataset through distributed normalization processing based on the raw multi-source biological data. The omics feature module performs autoencoder feature fusion based on the preprocessed biological dataset to generate a fused multi-omics feature vector. Distributed feature vector processing is then performed on the preprocessed biological dataset and the fused multi-omics feature vector to obtain the first multi-omics feature vector. The biological mining module generates a graph stitching probability matrix from the first multi-omics feature vector to obtain an edge probability matrix. Based on the edge probability matrix, the biological network mining results are obtained through feedback mining. The automatic analysis module processes the results of biological network mining through containerized scheduling decision-making to obtain task scheduling decisions, performs automated state analysis based on task scheduling decisions, and generates automated analysis results. The analysis and feedback module generates an analysis and feedback report based on the automated analysis results through verification and update feedback processing.

[0059] Example 3: Due to the insufficient proportion of driver genes caused by false positive interactions in existing technologies, the verification success rate is low, and there is a lack of confidence assessment and knowledge correction. As a result, the final target list contains a large number of false positives, leading to low accuracy of the results. Eliminating false positive targets consumes a lot of time and cost. This example also provides a computer-readable storage medium storing computer program instructions, which, when executed, implement the steps of any of the cloud computing-based biological data deep analysis methods mentioned above.

[0060] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code, including but not limited to disk storage, CD-ROM, optical storage, etc.

[0061] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A cloud computing-based method for deep analysis of biological data, characterized in that, The method involves the following steps: collecting raw multi-source biological data of the target organism, and obtaining a preprocessed biological dataset by performing distributed normalization processing on the raw multi-source biological data. Autoencoder feature fusion is performed on the preprocessed biological dataset to generate a fused multi-omics feature vector. Distributed feature vector processing is then performed on the preprocessed biological dataset and the fused multi-omics feature vector to obtain the first multi-omics feature vector. The graph splicing probability matrix is ​​generated from the first multi-omics feature vector to obtain the edge probability matrix. Based on the edge probability matrix, the biological network mining results are obtained through feedback mining. The process of processing the edge probability matrix through feedback mining includes: processing the edge probability matrix and the initial molecular graph through feedback iteration to obtain the enhanced node embedding and the first edge probability matrix. Key sub-confidence is evaluated based on the first-side probability matrix and enhanced node embedding, generating key sub-networks and network confidence; The network confidence, the first side probability matrix, the enhanced node embedding, and the first multi-omics feature vector are subjected to secondary external feedback processing to obtain the biological network mining results. Based on the results of biological network mining, a containerized scheduling decision is obtained through processing to obtain task scheduling decisions. Based on the task scheduling decisions, an automated state analysis is performed to generate automated analysis results. An analysis feedback report is obtained by verifying and updating the feedback based on the automated analysis results.

2. The method for deep analysis of biological data based on cloud computing according to claim 1, characterized in that, The process involves distributed normalization of raw multi-source biological data, including generating a unified cloud-based data format through distributed processing of the raw multi-source biological data. Cleaned biological data is obtained by performing cleaning and noise reduction based on a unified cloud data format. The cleaned biological data were normalized to obtain a preprocessed biological dataset.

3. The method for deep analysis of biological data based on cloud computing according to claim 1, characterized in that, Autoencoder feature fusion based on preprocessed biological datasets includes: processing the preprocessed biological datasets using an autoencoder to generate omics-specific coding vectors; Multi-head self-attention processing is performed based on omics-specific encoding vectors to generate an attention weight matrix; Cross-group feature fusion is performed on the attention weight matrix to generate a fused multi-omics feature vector.

4. The method for deep analysis of biological data based on cloud computing according to claim 1, characterized in that, Distributed feature vector processing is performed on the preprocessed biological dataset and the fused multi-omics feature vector, including: obtaining the fusion model parameters by distributed training based on the fused multi-omics feature vector and the preprocessed biological dataset; Feature importance is evaluated based on fusion model parameters to generate omics importance weights; The importance weights of omics and the fused multi-omics feature vectors are processed into multi-omics feature vectors to obtain the first multi-omics feature vector.

5. The method for deep analysis of biological data based on cloud computing according to claim 1, characterized in that, The graph stitching probability matrix generation for the first multi-omics feature vector includes: obtaining an initial molecular graph by graph stitching based on the first multi-omics feature vector; The node embedding representation is obtained by performing forward propagation of edge weights based on the initial molecular graph. The probability matrix of the initial molecular graph and node embedding representation is estimated to generate the edge probability matrix.

6. The method for deep analysis of biological data based on cloud computing according to claim 1, characterized in that, Based on the results of biological network mining, containerized scheduling decision-making is used for processing, including: generating containerized analysis pipelines based on the results of biological network mining through containerized orchestration. Based on containerized analysis pipelines, cloud resource awareness processing is used to obtain the resource status matrix; The resource state matrix and containerized analysis pipeline are processed for scheduling decisions to obtain task scheduling decisions.

7. The method for deep analysis of biological data based on cloud computing according to claim 1, characterized in that, Automated state analysis based on task scheduling decisions includes: processing state feedback based on task scheduling decisions and containerized analysis pipelines to obtain execution state vectors; The recovery scheduling instruction is obtained by performing backtracking checks based on the execution state vector. Automated analysis of execution state vectors and recovery scheduling instructions is performed to generate automated analysis results.

8. The method for deep analysis of biological data based on cloud computing according to claim 1, characterized in that, The process involves verifying and updating feedback based on automated analysis results, including: obtaining a verification report by performing bio-validation based on the automated analysis results. Based on the results of automated analysis, performance evaluation is conducted, and resource performance reports are generated. The verification report and resource performance report are updated and fed back to obtain an analysis feedback report.

9. A system applied to the cloud computing-based deep analysis method for biological data as described in any one of claims 1-8, characterized in that, The system includes: a preprocessing module, which collects raw multi-source biological data of the target organism, and obtains a preprocessed biological dataset by performing distributed normalization processing on the raw multi-source biological data; The omics feature module performs autoencoder feature fusion based on the preprocessed biological dataset to generate a fused multi-omics feature vector. Distributed feature vector processing is then performed on the preprocessed biological dataset and the fused multi-omics feature vector to obtain the first multi-omics feature vector. The biological mining module generates a graph stitching probability matrix from the first multi-omics feature vector to obtain an edge probability matrix. Based on the edge probability matrix, the biological network mining results are obtained through feedback mining. The automatic analysis module processes the results of biological network mining through containerized scheduling decision-making to obtain task scheduling decisions, performs automated state analysis based on task scheduling decisions, and generates automated analysis results. The analysis and feedback module generates an analysis and feedback report based on the automated analysis results through verification and update feedback processing.

10. A computer-readable storage medium having stored thereon computer program instructions that, when executed, implement the steps of the cloud computing-based biological data deep analysis method according to any one of claims 1 to 8.

Images