Bioinformatics data processing and analysis integrated system based on artificial intelligence

Abstract

The invention provides a bioinformatics data processing and analysis integrated system based on artificial intelligence, and relates to the technical field of bioinformatics and artificial intelligence crossing, and the system comprises a data collection module which is used for obtaining original data from a bioinformatics data source; according to the artificial intelligence bioinformatics data processing and analysis integrated system, a closed-loop control mechanism based on artificial intelligence is introduced into a bioinformatics data processing flow, so that collaborative linkage among data preprocessing, model modeling and result evaluation is realized. The system automatically generates feedback adjustment suggestions according to a comparison condition of a model analysis result and a preset performance standard, and dynamically adjusts a combination mode of a data preprocessing process and parameters and structure configuration of a prediction model, so that the problem that data processing and a model algorithm are mutually separated in a traditional process is avoided, and the efficiency is improved. And the analysis process is continuously and adaptively optimized along with the change of data characteristics and task requirements.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics and artificial intelligence, in particular to a bioinformatics data processing and analysis integrated system based on artificial intelligence. BACKGROUND

[0002] Current bioinformatics data processing mainly relies on scattered professional software tool chains. Researchers need to manually combine multiple independent tools such as sequence alignment, variant detection, and differential expression analysis to complete the process. Some solutions use workflow management systems to encapsulate fixed steps to achieve process standardization. At the same time, public and private biological databases provide a basis for data storage management. In addition, for specific tasks such as protein structure prediction and disease classification, several independent algorithm models based on machine learning frameworks have appeared.

[0003] However, the existing technology has a core defect that data processing, intelligent algorithms, and system platforms are mutually isolated, and cannot form a closed-loop optimization system. There is a lack of collaborative feedback mechanism between independent intelligent models and underlying data processing processes. Models can only passively receive pre-processed data in a fixed format, and model analysis results and performance indicators cannot automatically and intelligently guide the optimization and adjustment of standardized data preprocessing processes and feature engineering. This leads to poor adaptability and low iteration efficiency of the entire analysis process, which seriously restricts the systematic improvement of the actual effectiveness of artificial intelligence in bioinformatics. SUMMARY

[0004] To overcome the shortcomings of the prior art, the present application provides a bioinformatics data processing and analysis integrated system based on artificial intelligence, which solves the technical problem of how to build a model evaluation result to drive the closed-loop control of data preprocessing and prediction model parameter collaborative adjustment, and realize the adaptive linkage optimization between bioinformatics data processing processes and artificial intelligence models.

[0005] To achieve the above purpose, the present application realizes the following technical solutions: a bioinformatics data processing and analysis integrated system based on artificial intelligence, comprising: a data acquisition module for acquiring raw data from a bioinformatics data source.

[0006] An intelligent data preprocessing module is used for intelligent preprocessing of the raw data to generate preprocessed data, and the intelligent preprocessing adopts a configurable data preprocessing process.

[0007] An intelligent modeling module is used for constructing a prediction model, and the prediction model performs deep prediction learning on the preprocessed data to generate model analysis results.

[0008] An evaluation analysis module is used for multi-dimensional evaluation of the model analysis results to generate a comprehensive evaluation report.

[0009] An intelligent feedback mechanism module is configured to compare the comprehensive evaluation report with preset performance standards to generate a comparison result, and generate a feedback adjustment suggestion based on the comparison result.

[0010] A dynamic strategy adjustment module is configured to select a combination of the data preprocessing procedures and adjust parameters and architecture configurations of the prediction model based on the feedback adjustment suggestion, thereby forming a closed loop control.

[0011] Preferably, the bioinformatics data sources include genomic data, transcriptomic data, proteomic data, metabolomic data, clinical data, and public databases.

[0012] Preferably, the intelligent preprocessing includes format unification of the raw data to generate standardized intermediate data, quality dynamic evaluation of the standardized intermediate data to generate evaluation data, filtering of the standardized intermediate data that is lower than a preset quality indicator threshold, dimension reduction processing of the evaluation data to obtain the preprocessing data, and two-layer architecture for the dimension reduction processing, including a first layer and a second layer. The first layer uses a principal component analysis method for preliminary dimension reduction to reduce the feature dimension of the evaluation data to a preset intermediate dimension, and the second layer uses a variational autoencoder for deep representation learning to further compress the intermediate dimension to a target dimension.

[0013] Preferably, the data preprocessing procedures are each preprocessing step in the intelligent preprocessing as an independently callable processing unit, and the execution order of the preprocessing steps is combined based on the type of the bioinformatics data source and a preset flow mapping rule to form a data preprocessing procedure corresponding to the type of bioinformatics data.

[0014] Preferably, the construction of the prediction model is based on the type specificity of the bioinformatics data source, and the deep prediction learning includes inputting the preprocessing data into the prediction model, determining the bioinformatics data type corresponding to the preprocessing data based on the data type identification carried by the preprocessing data in the prediction model, extracting multi-scale features corresponding to the preprocessing data in the multi-scale feature fusion layer of the prediction model, and generating the model analysis result by weighting the multi-scale features, wherein the weighting calculates the correlation weight between different scale features.

[0015] Preferably, the multi-scale feature fusion layer adopts a hierarchical attention mechanism, the hierarchical attention mechanism includes a bottom layer attention and a high layer attention, the bottom layer attention calculates corresponding importance weights for different scale features within the same bioinformatics data type, the high layer attention calculates complementary weights for feature representations between different bioinformatics data types, and the multi-scale feature fusion layer generates the relevance weights by weighting the multi-scale features based on the importance weights and the complementary weights.

[0016] Preferably, the multi-dimensional evaluation is based on a preset evaluation index system to score the model analysis results in terms of computing performance, prediction accuracy and biological significance to generate evaluation index values, the evaluation index values are weighted and summed according to a preset weight ratio to obtain a total evaluation score, and the comprehensive evaluation report is formed based on the total evaluation score.

[0017] Preferably, the generation of the feedback adjustment suggestion includes comparing the scores of the comprehensive evaluation report with threshold values of preset performance standards item by item, identifying the evaluation dimensions that do not meet the standards, extracting intermediate data features associated with the evaluation dimensions that do not meet the standards from the process of generating the model analysis results, comparing the intermediate data features with the preset performance standards to obtain difference information, and integrating the difference information and the evaluation dimensions that do not meet the standards to generate the feedback adjustment suggestion.

[0018] Preferably, the selection of the data preprocessing process includes adjusting the enable state and execution order of each preprocessing step in the intelligent preprocessing process according to the difference information in the feedback adjustment suggestion, and adding the type of the bioinformatics data source to the prediction model according to the evaluation dimensions that do not meet the standards in the feedback adjustment suggestion.

[0019] Preferably, the closed loop control includes applying the adjusted data preprocessing process to the original data to generate new preprocessing data, inputting the new preprocessing data into the adjusted prediction model to generate new model analysis results, and outputting a new comprehensive evaluation report by the evaluation analysis module to drive the intelligent feedback mechanism module to generate a new feedback adjustment suggestion to form a closed loop.

[0020] The present application provides an artificial intelligence-based bioinformatics data processing and analysis integrated system, which has the following advantages: The application realizes the collaborative linkage among data preprocessing, model modeling and result evaluation by introducing an artificial intelligence-based closed-loop control mechanism in the bioinformatics data processing flow. The system automatically generates feedback adjustment suggestions according to the comparison between the model analysis results and the preset performance standards, and dynamically adjusts the combination mode of the data preprocessing flow and the parameters and structure configuration of the prediction model, avoiding the problem of mutual fragmentation of data processing and model algorithm in the traditional flow, so that the analysis flow continuously and adaptively optimizes with the changes of data characteristics and task requirements.

[0021] The artificial intelligence-based bioinformatics data processing and analysis integrated system improves the comprehensive performance of model analysis results in terms of prediction accuracy, computational efficiency and biological interpretability by integrating type-specific modeling of multi-source bioinformatics data, multi-scale feature fusion and multi-dimensional evaluation mechanism in a unified system. The system reduces manual intervention and repeated parameter adjustment process, shortens the analysis iteration cycle, enhances the stability and scalability under complex bioinformatics tasks, and is conducive to the large-scale landing of artificial intelligence technology in actual bioinformatics research and application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is a bioinformatics data processing and analysis integrated system structure schematic diagram; Figure 2 is an intelligent data preprocessing flowchart for implementing the application; Figure 3 is an intelligent modeling module structure schematic diagram for implementing the application; Figure 4 is an evaluation and feedback mechanism schematic diagram for implementing the application; Figure 5 is a dynamic strategy adjustment flowchart for implementing the application. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0024] As shown in Figures 1-5 , the artificial intelligence-based bioinformatics data processing and analysis integrated system provided by the embodiments of the application includes a data acquisition module for acquiring raw data from bioinformatics data sources. The bioinformatics data sources include genomic data, transcriptomic data, proteomic data, metabolomic data, clinical data and public databases.

[0025] The intelligent data preprocessing module is configured to preprocess the raw data to generate preprocessed data. The intelligent preprocessing adopts a configurable data preprocessing flow. The intelligent preprocessing includes: uniformly formatting the raw data to generate standardized intermediate data, dynamically evaluating the quality of the standardized intermediate data to generate evaluation data, filtering the standardized intermediate data that is lower than a preset quality indicator threshold, performing dimension reduction processing on the evaluation data to obtain preprocessed data, and adopting a two-layer architecture for the dimension reduction processing. The two-layer architecture includes a first layer and a second layer. The first layer adopts a principal component analysis method to perform preliminary dimension reduction, and reduces the feature dimension of the evaluation data to a preset intermediate dimension. The second layer adopts a variational autoencoder to perform deep representation learning, and further compresses the intermediate dimension to a target dimension. The data preprocessing flow is configured to combine the execution order of the preprocessing steps in the intelligent preprocessing as independently callable processing units according to the type of the bioinformatics data source and a preset flow mapping rule, and form a data preprocessing flow corresponding to the type of the bioinformatics data.

[0026] Raw data and format unification processing: The raw data is selected from breast cancer transcriptome sequencing data, which is obtained from the breast cancer project of the Cancer Genome Atlas project, and includes 1098 cases of transcriptome sequencing data and corresponding clinical information data.

[0027] The raw transcriptome data is stored in the form of a gene count matrix, with rows representing gene entries and columns representing sample numbers. The clinical data is stored with case numbers as indexes. The intelligent data preprocessing module performs format unification processing on data from different sources, including sample identification unification, gene identification unification, and numerical format unification processing.

[0028] At the sample level, the case number is used as the unique primary key. In the case of multiple sequencing samples for the same case, the sample with the highest total sequencing count is retained, and the remaining samples are discarded.

[0029] At the gene level, all gene names are uniformly mapped to Ensembl gene numbers, and entries that cannot be mapped are directly deleted.

[0030] In terms of numerical processing, the original gene count value is standardized and converted. The original count value of a gene in the jth sample is , and the total count of the sample is The standardized expression value of the gene is calculated as:

[0031] After the format unification processing is completed, a standardized intermediate data matrix is formed, with 1098 samples and 60483 gene entries.

[0032] Quality dynamic evaluation and threshold filtering: After obtaining the normalized intermediate data, the intelligent data preprocessing module dynamically evaluates the data quality and generates evaluation data. The quality evaluation is performed in both sample dimension and gene dimension.

[0033] The evaluation indicators in sample dimension include total sequencing count, zero value proportion, median expression value, and abnormal deviation degree. The abnormal deviation degree is calculated based on the principal component analysis result to measure the deviation of samples in low-dimensional space, which is used to identify obviously abnormal samples. The evaluation results show that the minimum value of sample total sequencing count is , the maximum value is , the median value is , and the median value of sample zero value proportion is 0.78.

[0034] The evaluation indicators in gene dimension include expression coverage and coefficient of variation. The expression coverage is defined as the proportion of samples with expression value not less than 1 in all samples; the coefficient of variation is used to measure the expression fluctuation degree of genes among samples.

[0035] In the quality filtering stage, the intelligent data preprocessing module filters the evaluation data using preset quality indicator thresholds.

[0036] The sample filtering rules include samples with total sequencing count less than , samples with zero value proportion higher than 0.93, and samples with abnormal deviation degree in the top 1% of the total samples.

[0037] The gene filtering rules include genes with expression coverage less than 20% and genes with coefficient of variation less than 0.05.

[0038] After the above filtering process, the number of samples is reduced from 1098 to 1062, and the number of gene entries is reduced from 60483 to 18962, obtaining the quality-filtered feature matrix.

[0039] Two-layer dimension reduction generates preprocessed data: After completing the quality filtering, the intelligent data preprocessing module performs dimension reduction on the evaluated data, and the dimension reduction process adopts a two-layer architecture.

[0040] The first layer is principal component analysis dimension reduction. After zero-mean and unit-variance standardization of the filtered feature matrix, principal component analysis method is used for preliminary dimension reduction, and 256 principal components are retained. The 256-dimensional feature explains 92.4% of the total variance information of the original data, forming the intermediate-dimensional representation data.

[0041] The second layer is a variational autoencoder dimension reduction. The variational autoencoder is a deep representation learning method based on a probabilistic generative model, the encoder network of which maps the 256-dimensional intermediate features to low-dimensional latent variables, and the decoder network of which is used to reconstruct the input features. The dimension of the latent variable is set to 32. After the model training is completed, the reconstruction mean square error on the validation set is 0.018, indicating that the low-dimensional representation can well preserve the original feature information.

[0042] The preprocessed data output by the intelligent data preprocessing module is a low-dimensional representation data of 1062 samples and 32-dimensional features, and the corresponding quality evaluation results, filtered records and dimension reduction parameter information are also output, which are used for deep prediction analysis of the subsequent intelligent modeling module.

[0043] The intelligent modeling module is used to build a prediction model, and the prediction model performs deep prediction learning on the preprocessed data to generate model analysis results. The construction of the prediction model is based on the type specificity of the bioinformatics data source, and the deep prediction learning includes inputting the preprocessed data into the prediction model, determining the bioinformatics data type corresponding to the preprocessed data according to the data type identification carried in the preprocessed data, extracting the multi-scale features corresponding to the preprocessed data in the multi-scale feature fusion layer of the prediction model, generating the model analysis results by weighting the multi-scale features, and calculating the correlation weight between different scale features by weighting. The multi-scale feature fusion layer adopts a hierarchical attention mechanism, which includes a bottom layer attention and a high layer attention. The bottom layer attention calculates the importance weight of different scale features within the same bioinformatics data type, and the high layer attention calculates the complementary weight of feature representation between different bioinformatics data types. The multi-scale feature fusion layer generates the correlation weight by weighting the multi-scale features based on the importance weight and the complementary weight.

[0044] The evaluation analysis module is used to perform multi-dimensional evaluation on the model analysis results to generate a comprehensive evaluation report. The multi-dimensional evaluation is based on a preset evaluation index system to score the model analysis results in terms of computing performance, prediction accuracy and biological significance to generate evaluation index values, to sum the evaluation index values by weighting according to a preset weight proportion to obtain an overall evaluation score, and to form a comprehensive evaluation report based on the overall evaluation score.

[0045] The input data of the evaluation analysis module includes: The prediction result data output by the intelligent modeling module includes the prediction category of each sample, the prediction probability of each category and the multi-scale feature fusion weight. The true label data corresponding to each sample is derived from the original clinical annotation field. The computing resource consumption data recorded during the model training and inference process, including single-round training time consumption, single-sample inference time consumption and model parameter size.

[0046] The evaluation index system is composed of three types of indexes: prediction performance indexes, calculation performance indexes and biological significance indexes. The weights of each type of index in the comprehensive evaluation are 0.5, 0.2 and 0.3 respectively. The above weight configuration is written into the evaluation configuration file and is fixed for use in the system.

[0047] Prediction performance evaluation and result data: The prediction performance evaluation is based on the test set data, and the number of test set samples is 160. The evaluation indexes include classification accuracy, macro-averaged precision, macro-averaged recall and macro-averaged F1 value.

[0048] Taking the classification accuracy as an example, the calculation formula is:

[0049] Wherein, N is the number of test set samples, is the model prediction category, is the true category.

[0050] The number of correct samples predicted by the model on the test set is 130, and the classification accuracy is:

[0051] The macro-averaged F1 value is obtained by calculating the F1 value of three types of samples respectively and taking the arithmetic mean, and the result is 0.79. The prediction performance evaluation results form the prediction performance evaluation table, and are important input for subsequent weighted calculation.

[0052] Calculation performance evaluation and result data: The calculation performance evaluation is used to measure the efficiency and resource consumption of the model in the actual running process. The evaluation indexes include the average time consumption of single round training, the time consumption of single sample inference and the model parameter size.

[0053] The model completes the training in a single graphics processing unit environment. The statistical results show that: The average time consumption of single round training is 42 seconds, and the total time consumption of 60 rounds of training is about 42 minutes. The time consumption of single sample inference is 3.8 milliseconds. The total amount of model parameters is about 2.1×10 6 parameters.

[0054] The calculation performance indexes are scored according to the preset interval. When the time consumption of single sample inference is less than 5 milliseconds and the model parameter size is less than 5×10 6 , the corresponding calculation performance score is recorded as full score 1.0. The calculation performance score result is 0.92.

[0055] Biological significance evaluation and result data: Biological significance evaluation is used to analyze the consistency between the model output features and known biological laws. The evaluation analysis module performs statistics on the potential features with higher weights in the multi-scale feature fusion layer, extracts the top 10% of potential feature dimensions in terms of contribution, and traces back to the gene contribution in the original transcriptome data.

[0056] The statistical results show that among the top 10% of potential features in terms of contribution, 17 features correspond to gene sets with clear annotations in the breast cancer-related pathway database, accounting for 68%. These are used as biological consistency indicators to input the evaluation model.

[0057] The biological significance score is calculated using a proportional mapping method. When the consistency ratio is higher than 60%, the score is not less than 0.8. In this example, the biological significance score is 0.84.

[0058] Comprehensive evaluation calculation and evaluation report generation: The evaluation analysis module weights and sums the three types of evaluation indicators according to the preset weights. The calculation formula of the comprehensive evaluation score S is:

[0059] Substituting the scores, the prediction performance score is 0.81, the calculation performance score is 0.92, and the biological significance score is 0.84. Therefore, the comprehensive evaluation score is:

[0060] The evaluation analysis module generates a comprehensive evaluation report. The report includes the original data, calculation results, weighted scores, and evaluation items that do not meet the preset threshold for each evaluation dimension, providing clear data basis for the subsequent intelligent feedback mechanism module.

[0061] Intelligent feedback mechanism module: used to compare the comprehensive evaluation report with the preset performance standard to generate a comparison result, and generate feedback adjustment suggestions based on the comparison result. The generation of feedback adjustment suggestions includes comparing the scores of the comprehensive evaluation report with the threshold of the preset performance standard, identifying the evaluation dimensions that do not meet the standard, extracting the intermediate data features associated with the evaluation dimensions that do not meet the standard from the process of generating model analysis results, comparing the intermediate data features with the preset performance standard to obtain difference information, and integrating the difference information and the evaluation dimensions that do not meet the standard to generate feedback adjustment suggestions.

[0062] Dynamic strategy adjustment module: used for adjusting the combination of data preprocessing procedures according to the feedback adjustment suggestion, adjusting the parameters and architecture configuration of the prediction model, forming a closed loop control. The selection of the data preprocessing procedure includes adjusting the enabled state and execution order of each preprocessing step in the intelligent preprocessing procedure according to the difference information in the feedback adjustment suggestion, and adding the type of bioinformatics data source to the prediction model according to the evaluation dimension that does not meet the standard in the feedback adjustment suggestion. The closed loop control includes applying the adjusted data preprocessing procedure to the original data to generate new preprocessing data, inputting the new preprocessing data into the adjusted prediction model to generate new model analysis results, and driving the intelligent feedback mechanism module to generate new feedback adjustment suggestions to form a closed loop.

[0063] Embodiment two This embodiment realizes effective classification of case classification and outputs quantifiable model analysis results by introducing data type-specific modeling and hierarchical attention-driven multi-scale feature fusion mechanism to perform deep prediction learning on the preprocessed transcriptome data.

[0064] 1. Input data and task definition The input data comes from the output of the intelligent data preprocessing module, the sample number is N=1062, and the low-dimensional feature vector dimension of each sample is d=32, denoted as the feature matrix: .

[0065] The prediction task is a case classification task, and the class label comes from the clinical annotation field matched with the sample number, with a total of 3 classes. The label distribution is: 423 cases of class I, 351 cases of class II, and 288 cases of class III. The label is represented as: .

[0066] The data set is divided by stratified sampling, with 70% for training set, 15% for validation set, and 15% for test set. The proportion difference of samples in each class in the three parts is controlled within ±2%.

[0067] 2. Data type-specific modeling and type identification processing The preprocessed data carries the data type identification, and the data type identification in this example is transcriptome data. After the model receives the input, the type identification is mapped to the type embedding vector, and the type embedding dimension is , and the type embedding is represented as: .

[0068] The type embedding vector is spliced with the sample feature to obtain the enhanced input:

[0069] wherein, is the 32-dimensional preprocessed feature of the i-th sample.

[0070] 3. Multi-scale feature extraction and multi-scale feature fusion layer 3.1 Multi-scale feature extraction For Three groups of scale branches are constructed to extract local scale features, medium scale features and global scale features respectively. The output dimensions of the three groups of branches are unified to 32 dimensions, denoted as:

[0071] The specific parameters of the three groups of branches are: Local scale branch: two layers of fully connected network, hidden layer dimension 64, output dimension 32. Medium scale branch: one layer of fully connected network, hidden layer dimension 128, output dimension 32. Global scale branch: one layer of fully connected network with residual connection, output dimension 32.

[0072] 3.2 Hierarchical attention mechanism The multi-scale feature fusion layer adopts hierarchical attention mechanism, including bottom attention and high-level attention.

[0073] The bottom attention is used to calculate the importance weight of different scale features in the same data type. Let the scale number be K=3, the bottom attention weight is:

[0074] where, is the attention score of the k-th scale branch to the sample iii.

[0075] In the validation set statistics, the three scale average weights are: local scale , medium scale , and global scale .

[0076] The high-level attention is used to calculate the complementarity weight between different data types. When the transcriptome data type is enabled, the complementarity weight of the high-level attention is fixed at 1.00, which is used to maintain interface consistency and facilitate subsequent expansion to multiple types of input.

[0077] 3.3 Correlation weight and weighted fusion The correlation weight of multi-scale features is the product of the bottom attention weight and the complementarity weight. The complementarity weight is , set =1, then the correlation weight of the k-th scale is:

[0078] The feature after weighted fusion is represented as:

[0079] 4. Depth prediction learning and output model analysis results 4.1 Classification output layer The fusion feature The input classifier obtains three-class probability output:

[0080] The model analysis results include class prediction and corresponding confidence max( ).

[0081] 4.2 Training objective function The cross-entropy loss function is used, the number of training set samples is n, and the loss is:

[0082] 4.3 Training parameters and convergence data The training rounds are 60 rounds, the batch size is 64, the initial learning rate is 0.001; When the validation set loss does not decrease for 6 consecutive rounds, the learning rate is decayed by 0.5 times once, and at most 3 times.

[0083] At the end of training, the validation set loss decreases from 1.08 in the first round to 0.62 in the 60th round. The test set classification accuracy is 0.81, and the macro average F1 value is 0.79. The model outputs the predicted class of each sample, the three-class probability vector and the three-scale weight of the fusion layer, which are used as the input of the subsequent evaluation analysis module.

[0084] Example three This embodiment quantitatively analyzes the comprehensive evaluation results, dynamically adjusts the data preprocessing process and prediction model parameter configuration, forms a closed-loop optimization strategy, and improves the model prediction performance and overall analysis effect.

[0085] 1. Input data and adjustment trigger condition The input data received by the dynamic strategy adjustment module includes: comprehensive evaluation score and corresponding prediction performance score, calculation performance score and biological significance score. The evaluation analysis module identifies the unqualified evaluation dimensions and corresponding difference information. The data preprocessing process configuration parameters and prediction model parameter configuration records used in the current round of operation.

[0086] The system presets the comprehensive evaluation score threshold as 0.85. When the comprehensive evaluation score is lower than the threshold, the dynamic strategy adjustment process is triggered. The comprehensive evaluation score of this round of evaluation results is 0.836, which is lower than the preset threshold, so the strategy adjustment is started.

[0087] 2. Information analysis based on evaluation difference The dynamic strategy adjustment module analyzes the evaluation dimensions that do not meet the standard. The prediction performance score is 0.81, which is lower than the prediction performance sub-threshold of 0.85. Both the performance score and the biological significance score are higher than the corresponding sub-thresholds.

[0088] The analysis of the prediction performance difference information shows that: In the test set, the recall rate of class II samples is 0.72, which is significantly lower than the recall rates of class I and class III samples. In the multi-scale feature fusion layer, the average weight of the medium-scale feature is 0.41, and the combined weight of the local-scale and global-scale features is 0.59, indicating that the medium-scale feature is dominant. The above difference information is recorded as the basis data for strategy adjustment.

[0089] 3. Dynamic adjustment of data preprocessing process To address the problem of insufficient prediction performance, the dynamic strategy adjustment module adjusts the data preprocessing process. The adjustment strategy includes refining the parameters in the dimension reduction stage.

[0090] In the original preprocessing process, the latent feature dimension of the second layer of dimension reduction is 32 dimensions. According to the prediction performance analysis results, the system increases the latent feature dimension to 48 dimensions to increase the feature expression ability required for distinguishing class II samples.

[0091] After adjusting, the dimension reduction process is performed again to obtain new preprocessed data. The number of samples remains 1062, and the feature dimension increases from 32 to 48. The reconstruction mean square error of the latent feature increases from 0.018 to 0.021 before and after adjustment, which is within the system's allowed range.

[0092] 4. Adjustment of prediction model parameters and structure After completing the adjustment of the data preprocessing process, the dynamic strategy adjustment module synchronously updates the prediction model configuration. In this embodiment, the main adjustment contents include: 4.1 Weight balance strategy of multi-scale feature fusion layer The initial weight constraint coefficient of the original medium-scale branch is reduced from 1.0 to 0.8, while the initial weight constraint coefficients of the local-scale and global-scale branches are increased from 1.0 to 1.1 to alleviate the problem of single-scale feature dominance.

[0093] 4.2 Class weight adjustment of classification loss function To address the problem of low recall rate of class II samples, a class weight coefficient is introduced in the loss function. Let the weights of the three classes be , , , then in this example we take: =1.0, =1.3, =1.0 To enhance the punishment of the model's prediction error on the type II samples.

[0094] 4.3 Training parameter fine-tuning The maximum number of training rounds is adjusted from 60 rounds to 70 rounds, and the initial learning rate is reduced from 0.001 to 0.0008 to improve the convergence stability of the model in the later training stage.

[0095] 5. Adjustment effect record and closed-loop output After the above strategy adjustment is completed, the dynamic strategy adjustment module writes the updated data preprocessing process parameters, prediction model parameter configuration and adjustment reason into the strategy version record table for subsequent evaluation and comparison.

[0096] After retraining and evaluation using the adjusted strategy, the test set classification accuracy is improved to 0.835, the type II sample recall rate is improved to 0.79, and the comprehensive evaluation score is improved to 0.862, exceeding the preset threshold of 0.85. The above results are recorded as the effectiveness data of this round of strategy adjustment.

[0097] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. An integrated system for bioinformatics data processing and analysis based on artificial intelligence, characterized in that, include: Data acquisition module: used to acquire raw data from bioinformatics data sources; Intelligent data preprocessing module: used to intelligently preprocess the raw data to generate preprocessed data, wherein the intelligent preprocessing adopts a configurable data preprocessing flow; Intelligent modeling module: used to build a prediction model, which performs deep predictive learning on the preprocessed data to generate model analysis results; Evaluation and Analysis Module: Used to perform multi-dimensional evaluation of the model analysis results and generate a comprehensive evaluation report; Intelligent feedback mechanism module: used to compare the comprehensive evaluation report with the preset performance standards to generate comparison results, and generate feedback adjustment suggestions based on the comparison results; Dynamic strategy adjustment module: used to select the combination of the data preprocessing process based on the feedback adjustment suggestions, and adjust the parameters and architecture configuration of the prediction model to form closed-loop control.

2. The integrated system for bioinformatics data processing and analysis based on artificial intelligence according to claim 1, characterized in that: The bioinformatics data sources include genomic data, transcriptomic data, proteomic data, metabolomic data, clinical data, and public databases.

3. The integrated system for bioinformatics data processing and analysis based on artificial intelligence according to claim 1, characterized in that: The intelligent preprocessing includes unifying the format of the original data to generate standardized intermediate data, performing dynamic quality assessment on the standardized intermediate data to generate assessment data, wherein the dynamic quality assessment involves filtering the standardized intermediate data that is below a preset quality index threshold, and performing dimensionality reduction processing on the assessment data to obtain the preprocessed data.

4. The integrated system for bioinformatics data processing and analysis based on artificial intelligence according to claim 1, characterized in that: The data preprocessing process involves treating each preprocessing step in the intelligent preprocessing as an independently callable processing unit, and combining the execution order of the preprocessing steps according to the type of the bioinformatics data source and preset process mapping rules to form a data preprocessing process corresponding to the bioinformatics data type.

5. The integrated system for bioinformatics data processing and analysis based on artificial intelligence according to claim 1, characterized in that: The construction of the prediction model is based on the type specificity of the bioinformatics data source. The deep prediction learning includes inputting the preprocessed data into the prediction model, the prediction model determining the bioinformatics data type corresponding to the preprocessed data according to the data type identifier carried in the preprocessed data, extracting the multi-scale features corresponding to the preprocessed data in the multi-scale feature fusion layer of the prediction model, performing weighted processing on the multi-scale features to generate the model analysis results, and calculating the correlation weights between features of different scales in the weighted processing.

6. The integrated system for bioinformatics data processing and analysis based on artificial intelligence according to claim 5, characterized in that: The multi-scale feature fusion layer adopts a hierarchical attention mechanism, which includes bottom-level attention and high-level attention. The bottom-level attention calculates the corresponding importance weights for features at different scales within the same bioinformatics data type. The high-level attention calculates the complementarity weights for feature representations across different bioinformatics data types. Based on the importance weights and the complementarity weights, the multi-scale feature fusion layer generates the relevance weights by weighting the multi-scale features.

7. The integrated system for bioinformatics data processing and analysis based on artificial intelligence according to claim 1, characterized in that: The multi-dimensional assessment is based on a preset evaluation index system. The model analysis results are scored in terms of computational performance, prediction accuracy, and biological significance to generate evaluation index values. The evaluation index values ​​are weighted and summed according to preset weight ratios to obtain an overall evaluation score. The comprehensive evaluation report is formed based on the overall evaluation score.

8. The integrated system for bioinformatics data processing and analysis based on artificial intelligence according to claim 1, characterized in that: The generation of the feedback adjustment suggestions includes comparing the scores of the comprehensive evaluation report with the thresholds of the preset performance standards item by item to identify the evaluation dimensions that have not met the standards, extracting intermediate data features associated with the evaluation dimensions that have not met the standards from the process of generating the model analysis results, comparing the intermediate data features with the preset performance standards to obtain the difference information, and integrating the difference information and the evaluation dimensions that have not met the standards to generate the feedback adjustment suggestions.

9. The integrated system for bioinformatics data processing and analysis based on artificial intelligence according to claim 1, characterized in that: The selection of the data preprocessing process includes adjusting the activation status and execution order of each preprocessing step in the intelligent preprocessing process based on the difference information in the feedback adjustment suggestions, and adding the type of bioinformatics data source to the prediction model based on the evaluation dimensions that are not met in the feedback adjustment suggestions.

10. The integrated system for bioinformatics data processing and analysis based on artificial intelligence according to claim 1, characterized in that: The closed-loop control includes applying the adjusted data preprocessing process to the original data to generate new preprocessed data, inputting the new preprocessed data into the adjusted prediction model to generate new model analysis results, and the new comprehensive evaluation report output by the evaluation and analysis module driving the intelligent feedback mechanism module to generate new feedback adjustment suggestions to form a closed loop.

Images